{"id":183,"date":"2026-04-27T23:59:03","date_gmt":"2026-04-27T15:59:03","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/27\/llm-fine-tuning-from-full-to-lora-and-qlora\/"},"modified":"2026-04-27T23:59:03","modified_gmt":"2026-04-27T15:59:03","slug":"llm-fine-tuning-from-full-to-lora-and-qlora","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/27\/llm-fine-tuning-from-full-to-lora-and-qlora\/","title":{"rendered":"\u5927\u6a21\u578b\u5fae\u8c03\u5165\u95e8\uff1a\u4ece\u5168\u91cf\u5fae\u8c03\u5230 LoRA \u548c QLoRA \u2014\u2014 \u4e00\u4e2a\u5de5\u7a0b\u5e08\u7684\u5b9e\u6218\u7b14\u8bb0"},"content":{"rendered":"

\u6211\u7b2c\u4e00\u6b21\u5c1d\u8bd5\u5fae\u8c03 LLaMA-7B \u7684\u65f6\u5019\uff0c\u4ee5\u4e3a\u53ea\u8981\u628a\u6570\u636e\u5582\u8fdb\u53bb\u3001\u8c03\u4e2a\u5b66\u4e60\u7387\uff0c\u5c31\u80fd\u8ba9\u6a21\u578b\u5b66\u4f1a\u6211\u7684\u4e1a\u52a1\u95ee\u7b54\u98ce\u683c\u3002\u7ed3\u679c\uff0c\u6211\u7684 24GB \u663e\u5b58\u7684 RTX 3090 \u76f4\u63a5\u7206\u4e86\uff0c\u8fde\u52a0\u8f7d\u6a21\u578b\u90fd\u5931\u8d25\u3002\u90a3\u4e00\u523b\u6211\u624d\u660e\u767d\uff1a\u5927\u6a21\u578b\u4e0d\u662f\u73a9\u5177\uff0c\u662f\u9700\u8981\u656c\u754f\u7684\u5de8\u517d\u3002<\/p>\n

\u90a3\u65f6\u5019\u6211\u8fd8\u4e0d\u77e5\u9053 LoRA \u662f\u4ec0\u4e48\uff0c\u53ea\u77e5\u9053\u5168\u91cf\u5fae\u8c03\uff08Full Fine-tuning\uff09\u8981\u5b58\u4e0b\u6574\u4e2a\u6a21\u578b\u7684\u68af\u5ea6\u548c\u4f18\u5316\u5668\u72b6\u6001\u2014\u20147B \u53c2\u6570\uff0c\u6bcf\u4e2a\u53c2\u6570 16 \u4f4d\uff0c\u5149\u662f\u4f18\u5316\u5668\u72b6\u6001\u5c31\u8981 56GB \u5185\u5b58\uff0c\u66f4\u522b\u63d0\u4e2d\u95f4\u6fc0\u6d3b\u503c\u4e86\u3002\u6211\u751a\u81f3\u6000\u7591\uff0c\u662f\u4e0d\u662f\u53ea\u6709\u5927\u5382\u624d\u80fd\u73a9\u5f97\u8d77\u5927\u6a21\u578b\uff1f<\/p>\n

\u76f4\u5230\u6211\u8bfb\u5230 LoRA \u7684\u8bba\u6587\uff0c\u624d\u50cf\u88ab\u4e00\u9053\u95ea\u7535\u5288\u4e2d\uff1a\u539f\u6765\u6211\u4eec\u4e0d\u9700\u8981\u52a8\u6a21\u578b\u7684\u2018\u5927\u8111\u2019\uff0c\u53ea\u9700\u8981\u5728\u5b83\u7684\u2018\u795e\u7ecf\u901a\u8def\u2019\u4e0a\u52a0\u51e0\u4e2a\u5c0f\u63d2\u4ef6\uff0c\u5c31\u80fd\u8ba9\u5b83\u5b66\u4f1a\u65b0\u6280\u80fd\u3002<\/p>\n

—<\/p>\n

\n\"Personal
Personal design algorithm\u2026 I\u2019ve been trying to envision how a machine learning\/neural network can aid a designer\u2014not design things for you, but help figure out what your design should be; an idea expander; a fast sketching partner\u2026 anyway, that\u2019s been mak – bjornmeansbear, by-sa<\/figcaption><\/figure>\n

\u4e3a\u4ec0\u4e48\u6211\u5f00\u59cb\u5173\u6ce8\u5927\u6a21\u578b\u5fae\u8c03\uff1f<\/h2>\n

\u6211\u662f\u4e00\u4e2a\u505a\u4f01\u4e1a\u7ea7\u5ba2\u670d\u5bf9\u8bdd\u7cfb\u7edf\u7684\u5de5\u7a0b\u5e08\u3002\u6211\u4eec\u60f3\u8ba9\u6a21\u578b\u7406\u89e3\u516c\u53f8\u5185\u90e8\u7684\u672f\u8bed\u3001\u6d41\u7a0b\u3001\u5de5\u5355\u7cfb\u7edf\uff0c\u800c\u4e0d\u662f\u8ba9\u5b83\u7ee7\u7eed\u5728\u4e92\u8054\u7f51\u4e0a\u2018\u80e1\u8bf4\u516b\u9053\u2019\u3002\u4f46\u516c\u53f8\u6ca1\u6709 10 \u5757 A100\uff0c\u4e5f\u6ca1\u6709 100 \u4e07\u9884\u7b97\u4e70 API\u3002\u6211\u4eec\u53ea\u80fd\u7528\u5f00\u6e90\u6a21\u578b\uff0c\u81ea\u5df1\u5fae\u8c03\u3002<\/p>\n

\u5168\u91cf\u5fae\u8c03\u7684\u95e8\u69db\u592a\u9ad8\u4e86\u3002\u6211\u8bd5\u8fc7\u7528 Hugging Face \u7684 transformers<\/code> \u5e93\u8dd1\u4e00\u4e2a 7B \u6a21\u578b\uff0c\u54ea\u6015\u53ea\u8bad\u7ec3 1 \u4e2a epoch\uff0c\u663e\u5b58\u5c31\u98d9\u5230 28GB\u3002\u66f4\u522b\u63d0\u8981\u4fdd\u5b58\u591a\u4e2a checkpoint\u3001\u505a\u9a8c\u8bc1\u3001\u8c03\u53c2\u2014\u2014\u6839\u672c\u8dd1\u4e0d\u8d77\u6765\u3002<\/p>\n

\u6211\u5f00\u59cb\u5bfb\u627e\u66ff\u4ee3\u65b9\u6848\u3002LoRA \u662f\u7b2c\u4e00\u4e2a\u8ba9\u6211\u89c9\u5f97\u2018\u8fd9\u73a9\u610f\u513f\u9760\u8c31\u2019\u7684\u6280\u672f\u3002\u5b83\u4e0d\u6539\u53d8\u539f\u6a21\u578b\uff0c\u53ea\u5728\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u5c42\u63d2\u5165\u4f4e\u79e9\u77e9\u9635\uff0c\u8bad\u7ec3\u65f6\u53ea\u66f4\u65b0\u8fd9\u4e9b\u2018\u5c0f\u63d2\u4ef6\u2019\u3002\u6211\u7528 PEFT \u5e93\uff08Parameter-Efficient Fine-Tuning\uff09\u628a LoRA \u52a0\u5230 LLaMA \u4e0a\uff0c\u8bad\u7ec3\u53c2\u6570\u4ece 70 \u4ebf\u964d\u5230 100 \u4e07\uff0c\u663e\u5b58\u5360\u7528\u4ece 28GB \u964d\u5230 8GB\u3002\u90a3\u4e00\u523b\uff0c\u6211\u611f\u89c9\u81ea\u5df1\u4e0d\u662f\u5728\u2018\u8bad\u7ec3\u6a21\u578b\u2019\uff0c\u800c\u662f\u5728\u2018\u7ed9\u6a21\u578b\u7a7f\u4e00\u4ef6\u8f7b\u4fbf\u7684\u5916\u8863\u2019\u3002<\/p>\n

—<\/p>\n

LoRA \u7684\u6838\u5fc3\u673a\u5236\uff1a\u4f4e\u79e9\u77e9\u9635\u7684\u9b54\u6cd5<\/h2>\n

LoRA \u7684\u6838\u5fc3\u601d\u60f3\uff0c\u5176\u5b9e\u975e\u5e38\u4f18\u96c5\uff1a<\/p>\n

\u6211\u4eec\u5047\u8bbe\uff0c\u6a21\u578b\u5728\u9002\u5e94\u65b0\u4efb\u52a1\u65f6\uff0c\u53c2\u6570\u7684\u2018\u53d8\u5316\u2019\uff08\u0394W\uff09\u662f\u4f4e\u79e9\u7684\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u867d\u7136\u539f\u59cb\u6743\u91cd\u77e9\u9635 W \u662f 4096\u00d74096 \u7684\u5927\u77e9\u9635\uff0c\u4f46\u5b83\u7684\u2018\u8c03\u6574\u91cf\u2019 \u0394W \u53ef\u4ee5\u7528\u4e24\u4e2a\u5c0f\u77e9\u9635 A \u548c B \u7684\u4e58\u79ef\u6765\u8fd1\u4f3c\uff1a<\/p>\n

\\Delta W = A \\cdot B<\/span><\/p>\n

\u5176\u4e2d A \u662f 4096\u00d7r\uff0cB \u662f r\u00d74096\uff0cr \u662f\u4e00\u4e2a\u5f88\u5c0f\u7684\u6570\uff0c\u6bd4\u5982 8 \u6216 16\u3002<\/p>\n

\u8fd9\u6837\uff0c\u539f\u672c\u9700\u8981\u66f4\u65b0 4096\u00d74096 = 1678 \u4e07\u4e2a\u53c2\u6570\uff0c\u73b0\u5728\u53ea\u9700\u8981\u66f4\u65b0 4096\u00d78 + 8\u00d74096 = 65,536 \u4e2a\u53c2\u6570\u2014\u2014\u51cf\u5c11\u4e86 256 \u500d\uff01<\/p>\n

\u66f4\u5999\u7684\u662f\uff0c\u63a8\u7406\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u628a \u0394W \u5408\u5e76\u56de\u539f\u6743\u91cd\uff1a<\/p>\n

W_{\\text{new}} = W + A \\cdot B<\/span><\/p>\n

\u6240\u4ee5\u63a8\u7406\u901f\u5ea6\u548c\u539f\u6a21\u578b\u4e00\u6a21\u4e00\u6837\uff0c\u6ca1\u6709\u989d\u5916\u5ef6\u8fdf\u2014\u2014\u8fd9\u548c Adapter \u90a3\u79cd\u52a0\u4e2a\u5c0f\u578b\u7f51\u7edc\u7684\u65b9\u6848\u5b8c\u5168\u4e0d\u540c\uff0c\u540e\u8005\u6bcf\u6b21\u90fd\u8981\u591a\u8dd1\u4e00\u6b21\u524d\u5411\u4f20\u64ad\u3002<\/p>\n

\u6211\u7b2c\u4e00\u6b21\u5b9e\u73b0\u65f6\uff0c\u4ee5\u4e3a\u53ea\u8981\u628a A \u548c B \u521d\u59cb\u5316\u4e3a\u968f\u673a\u9ad8\u65af\u5206\u5e03\u5c31\u884c\u3002\u7ed3\u679c\u6a21\u578b\u5b8c\u5168\u5b66\u4e0d\u4f1a\u3002\u540e\u6765\u624d\u77e5\u9053\uff0cB \u8981\u521d\u59cb\u5316\u4e3a\u96f6\uff0cA \u8981\u521d\u59cb\u5316\u4e3a\u9ad8\u65af\uff0c\u8fd9\u6837\u521d\u59cb\u8f93\u51fa\u4e0d\u53d8\uff0c\u8bad\u7ec3\u624d\u7a33\u5b9a\u3002\u8fd9\u4e2a\u7ec6\u8282\uff0c\u8bba\u6587\u91cc\u53ea\u7528\u4e00\u53e5\u8bdd\u5e26\u8fc7\uff0c\u4f46\u5bf9\u5de5\u7a0b\u5b9e\u73b0\u81f3\u5173\u91cd\u8981\u3002<\/p>\n

—<\/p>\n

QLoRA\uff1a\u628a\u6a21\u578b\u2018\u538b\u7f29\u2019\u5230 4 \u4f4d\uff0c\u8fd8\u80fd\u5fae\u8c03\uff1f<\/h2>\n

LoRA \u5df2\u7ecf\u5f88\u8f7b\u4e86\uff0c\u4f46\u5f53\u6211\u5c1d\u8bd5\u5fae\u8c03 13B \u6216 33B \u6a21\u578b\u65f6\uff0c\u8fd8\u662f\u5361\u5728\u52a0\u8f7d\u9636\u6bb5\u2014\u2014\u5149\u662f\u628a\u6a21\u578b\u6743\u91cd\u4ece\u786c\u76d8\u8bfb\u8fdb\u663e\u5b58\uff0c\u5c31\u8981 26GB\uff0816 \u4f4d\uff09\u751a\u81f3 52GB\uff0832 \u4f4d\uff09\u3002\u6211\u7684 48GB A6000 \u4e5f\u625b\u4e0d\u4f4f\u3002<\/p>\n

\u76f4\u5230 QLoRA \u51fa\u73b0\u3002<\/p>\n

QLoRA \u7684\u4f5c\u8005\u4eec\u5e72\u4e86\u4e00\u4ef6\u2018\u75af\u72c2\u2019\u7684\u4e8b\uff1a\u4ed6\u4eec\u628a\u6574\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\u91cf\u5316\u5230 4 \u4f4d\uff08NormalFloat-4\uff0c\u7b80\u79f0 NF4\uff09\uff0c\u7136\u540e\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u4ecd\u7136\u7528 16 \u4f4d\u7cbe\u5ea6\u8ba1\u7b97\u68af\u5ea6\u3002\u542c\u8d77\u6765\u50cf\u9b54\u672f\uff0c\u4f46\u5b83\u662f\u771f\u7684\u3002<\/p>\n

\u4ed6\u4eec\u505a\u4e86\u4e09\u4ef6\u5173\u952e\u521b\u65b0\uff1a<\/p>\n

    \n
  • NF4 \u91cf\u5316<\/strong>\uff1a\u4e0d\u662f\u7528\u666e\u901a\u7684 int4\uff0c\u800c\u662f\u7528\u4e00\u79cd\u4e13\u95e8\u4e3a\u2018\u6b63\u6001\u5206\u5e03\u6743\u91cd\u2019\u8bbe\u8ba1\u7684 4 \u4f4d\u6d6e\u70b9\u683c\u5f0f\uff0c\u6bd4\u6807\u51c6 int4 \u66f4\u9002\u5408 LLM \u7684\u6743\u91cd\u5206\u5e03\uff0c\u4fe1\u606f\u635f\u5931\u66f4\u5c0f\u3002<\/li>\n
  • \u53cc\u91cf\u5316<\/strong>\uff1a\u8fde\u91cf\u5316\u65f6\u7528\u7684\u7f29\u653e\u56e0\u5b50\uff08scale\uff09\u548c\u504f\u79fb\uff08offset\uff09\u4e5f\u518d\u505a\u4e00\u6b21\u91cf\u5316\uff01\u8fd9\u542c\u8d77\u6765\u50cf\u2018\u538b\u7f29\u518d\u538b\u7f29\u2019\uff0c\u4f46\u5b9e\u6d4b\u80fd\u518d\u7701 10% \u5185\u5b58\u3002<\/li>\n
  • \u5206\u9875\u4f18\u5316\u5668<\/strong>\uff1a\u7528\u7c7b\u4f3c\u64cd\u4f5c\u7cfb\u7edf\u7684\u865a\u62df\u5185\u5b58\u673a\u5236\uff0c\u628a\u4f18\u5316\u5668\u72b6\u6001\u5206\u9875\u5b58\u50a8\u5728 CPU \u5185\u5b58\u548c\u663e\u5b58\u4e4b\u95f4\uff0c\u907f\u514d\u4e00\u6b21\u6027\u7206\u663e\u5b58\u3002<\/li>\n<\/ul>\n

    \u7ed3\u679c\uff1f\u4ed6\u4eec\u7528\u4e00\u5757 48GB \u7684 A6000\uff0c\u5fae\u8c03\u4e86 65B \u7684 LLaMA \u6a21\u578b\uff0c\u6027\u80fd\u8fbe\u5230 ChatGPT \u7684 99.3%\uff01<\/p>\n

    \u6211\u4eb2\u81ea\u8bd5\u4e86 QLoRA\u3002\u7528 bitsandbytes<\/code> \u5e93\u52a0\u8f7d 4-bit LLaMA-7B\uff0c\u663e\u5b58\u5360\u7528\u53ea\u6709 4.5GB\u3002\u52a0\u4e0a LoRA \u540e\uff0c\u8bad\u7ec3\u53c2\u6570\u8fd8\u662f 100 \u4e07\uff0c\u4f46\u6574\u4e2a\u6d41\u7a0b\u8dd1\u5f97\u6bd4 LoRA \u5728 7B \u4e0a\u8fd8\u5feb\u2014\u2014\u56e0\u4e3a\u6a21\u578b\u672c\u8eab\u5c0f\u4e86\uff01<\/p>\n

    \u8fd9\u5f7b\u5e95\u98a0\u8986\u4e86\u6211\u7684\u8ba4\u77e5\uff1a\u539f\u6765\u2018\u5927\u6a21\u578b\u2019\u4e0d\u662f\u2018\u5927\u663e\u5b58\u2019\u7684\u4ee3\u540d\u8bcd\uff0c\u800c\u662f\u2018\u5927\u80fd\u529b\u2019\u7684\u4ee3\u540d\u8bcd\u3002\u6211\u4eec\u4e0d\u9700\u8981\u628a\u6574\u4e2a\u5de8\u517d\u642c\u8fdb\u663e\u5b58\uff0c\u53ea\u9700\u8981\u8ba9\u5b83\u2018\u534a\u7761\u534a\u9192\u2019\uff0c\u7136\u540e\u8f7b\u8f7b\u63a8\u5b83\u4e00\u628a\u3002<\/p>\n

    —<\/p>\n

    \u5de5\u7a0b\u5b9e\u8df5\u4e2d\u7684\u8840\u6cea\u6559\u8bad<\/h2>\n

    \u7eb8\u4e0a\u5f97\u6765\u7ec8\u89c9\u6d45\u3002\u6211\u5728\u751f\u4ea7\u73af\u5883\u8e29\u8fc7\u7684\u5751\uff0c\u6bd4\u8bba\u6587\u91cc\u5199\u7684\u591a\u5341\u500d\u3002<\/p>\n

    1. LoRA \u7684 rank \u4e0d\u662f\u8d8a\u5927\u8d8a\u597d<\/h3>\n

    \u6211\u4e00\u5f00\u59cb\u4ee5\u4e3a r=64 \u4f1a\u6bd4 r=8 \u597d\uff0c\u7ed3\u679c\u8bad\u7ec3\u4e86\u4e24\u5929\uff0c\u9a8c\u8bc1\u96c6\u6548\u679c\u53cd\u800c\u4e0b\u964d\u3002\u540e\u6765\u53d1\u73b0\uff1ar \u592a\u5927\uff0c\u6a21\u578b\u5f00\u59cb\u8fc7\u62df\u5408\u3002\u5728\u5c0f\u6570\u636e\u96c6\u4e0a\uff0cr=8~16 \u6700\u7a33\u5b9a\u3002\u6211\u73b0\u5728\u7684\u7ecf\u9a8c\u6cd5\u5219\u662f\uff1a\u6570\u636e\u91cf < 10K \u6761\uff0c\u7528 r=8\uff1b10K~50K\uff0c\u7528 r=16\uff1b\u8d85\u8fc7 100K\uff0c\u518d\u8003\u8651 r=32\u3002<\/p>\n

    2. \u5b66\u4e60\u7387\u8981\u8c03\u5f97\u6bd4\u5168\u91cf\u5fae\u8c03\u9ad8<\/h3>\n

    LoRA \u7684\u53c2\u6570\u5c11\uff0c\u4f46\u6bcf\u4e2a\u53c2\u6570\u7684\u66f4\u65b0\u5e45\u5ea6\u8981\u66f4\u5927\u3002\u6211\u8bd5\u8fc7\u7528\u5168\u91cf\u5fae\u8c03\u7684 2e-5\uff0c\u7ed3\u679c\u6a21\u578b\u7eb9\u4e1d\u4e0d\u52a8\u3002\u540e\u6765\u6539\u6210 5e-4~1e-3\uff0c\u624d\u5f00\u59cb\u6709\u6548\u5b66\u4e60\u3002\u8fd9\u548c\u76f4\u89c9\u76f8\u53cd\u2014\u2014\u53c2\u6570\u5c11\uff0c\u53cd\u800c\u8981\u2018\u731b\u836f\u2019\u3002<\/p>\n

    3. \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u4e0d\u662f\u4e07\u80fd\u7684<\/h3>\n

    QLoRA \u7528\u7684\u662f 4-bit \u6a21\u578b + 16-bit \u68af\u5ea6\uff0c\u4f46\u5982\u679c\u4f60\u4e0d\u5c0f\u5fc3\u628a\u4f18\u5316\u5668\u4e5f\u8bbe\u6210 8-bit\uff0c\u68af\u5ea6\u4f1a\u6ea2\u51fa\u3002\u6211\u66fe\u56e0\u4e3a\u7528\u4e86 AdamW8bit<\/code> \u800c\u5bfc\u81f4\u635f\u5931\u7206\u70b8\uff0c\u6700\u540e\u53d1\u73b0\u5fc5\u987b\u7528 AdamW<\/code>\uff0816-bit\uff09\u914d\u5408 bitsandbytes<\/code> \u7684 4-bit \u6a21\u578b\u3002<\/p>\n

    4. \u6a21\u578b\u5408\u5e76\u8981\u5c0f\u5fc3<\/h3>\n

    \u8bad\u7ec3\u5b8c\u540e\uff0c\u5f88\u591a\u4eba\u60f3\u628a LoRA \u6743\u91cd\u5408\u5e76\u56de\u539f\u6a21\u578b\uff0c\u65b9\u4fbf\u90e8\u7f72\u3002\u4f46\u5982\u679c\u4f60\u7528\u7684\u662f QLoRA\uff0c\u5408\u5e76\u524d\u5fc5\u987b\u5148\u2018\u53cd\u91cf\u5316\u2019\u56de 16 \u4f4d\uff0c\u5426\u5219\u5408\u5e76\u540e\u7684\u6a21\u578b\u4f1a\u4e25\u91cd\u5931\u771f\u3002\u6211\u7b2c\u4e00\u6b21\u5408\u5e76\u540e\uff0c\u6a21\u578b\u8f93\u51fa\u5168\u662f\u4e71\u7801\uff0c\u82b1\u4e86\u4e24\u5929\u624d\u641e\u61c2\u662f\u91cf\u5316\u8bef\u5dee\u7d2f\u79ef\u5bfc\u81f4\u7684\u3002<\/p>\n

    5. \u4e0d\u8981\u8ff7\u4fe1\u2018SOTA\u2019<\/h3>\n

    QLoRA \u8bba\u6587\u8bf4 Guanaco \u5728 Vicuna \u57fa\u51c6\u4e0a\u63a5\u8fd1 ChatGPT\u3002\u4f46\u6211\u5728\u81ea\u5df1\u7684\u4e1a\u52a1\u6570\u636e\u4e0a\u8bd5\u4e86\uff0c\u6548\u679c\u8fd8\u4e0d\u5982\u4e00\u4e2a\u7cbe\u5fc3\u8bbe\u8ba1\u7684 Prompt + RAG\u3002\u5927\u6a21\u578b\u4e0d\u662f\u94f6\u5f39\uff0c\u5fae\u8c03\u4e5f\u4e0d\u662f\u4e07\u80fd\u836f\u3002\u6709\u65f6\u5019\uff0c\u52a0\u4e2a\u77e5\u8bc6\u5e93\uff0c\u6bd4\u5fae\u8c03 10 \u5c0f\u65f6\u66f4\u6709\u6548\u3002<\/p>\n

    —<\/p>\n

    \u6211\u7684\u9636\u6bb5\u6027\u7406\u89e3\uff1a\u5fae\u8c03\u7684\u672c\u8d28\u662f\u2018\u5f15\u5bfc\u2019\uff0c\u4e0d\u662f\u2018\u91cd\u5199\u2019<\/h2>\n

    \u7ecf\u8fc7\u534a\u5e74\u7684\u5b9e\u8df5\uff0c\u6211\u5bf9\u5927\u6a21\u578b\u5fae\u8c03\u7684\u7406\u89e3\u53d1\u751f\u4e86\u6839\u672c\u8f6c\u53d8\u3002<\/p>\n

    \u4ee5\u524d\u6211\u4ee5\u4e3a\uff0c\u5fae\u8c03\u5c31\u662f\u2018\u91cd\u65b0\u6559\u6a21\u578b\u2019\uff0c\u8ba9\u5b83\u5fd8\u8bb0\u65e7\u77e5\u8bc6\uff0c\u8bb0\u4f4f\u65b0\u77e5\u8bc6\u3002\u4f46\u73b0\u5728\u6211\u660e\u767d\uff1a\u5fae\u8c03\u7684\u672c\u8d28\u662f\u2018\u5f15\u5bfc\u2019\u3002<\/p>\n

    \u5927\u6a21\u578b\u5df2\u7ecf\u62e5\u6709\u6d77\u91cf\u7684\u901a\u7528\u8bed\u8a00\u80fd\u529b\uff0c\u5b83\u4e0d\u662f\u4e00\u5f20\u767d\u7eb8\uff0c\u800c\u662f\u4e00\u4e2a\u62e5\u6709\u767e\u79d1\u5168\u4e66\u7684\u5927\u8111\u3002\u6211\u4eec\u4e0d\u9700\u8981\u91cd\u5199\u5b83\uff0c\u53ea\u9700\u8981\u5728\u5b83\u601d\u8003\u65f6\uff0c\u8f7b\u8f7b\u62c9\u4e00\u4e0b\u5b83\u7684\u8863\u89d2\uff0c\u544a\u8bc9\u5b83\uff1a\u2018\u563f\uff0c\u8fd9\u6b21\u8bf7\u7528\u6211\u4eec\u516c\u53f8\u7684\u672f\u8bed\u6765\u56de\u7b54\u3002\u2019<\/p>\n

    LoRA \u548c QLoRA \u7684\u4f1f\u5927\u4e4b\u5904\uff0c\u4e0d\u5728\u4e8e\u5b83\u4eec\u2018\u8282\u7701\u4e86\u663e\u5b58\u2019\uff0c\u800c\u5728\u4e8e\u5b83\u4eec\u5c0a\u91cd\u4e86\u6a21\u578b\u7684\u2018\u9884\u8bad\u7ec3\u77e5\u8bc6\u2019\u3002\u5b83\u4eec\u4e0d\u7834\u574f\uff0c\u53ea\u589e\u5f3a\uff1b\u4e0d\u8986\u76d6\uff0c\u53ea\u6ce8\u5165\u3002<\/p>\n

    \u6211\u73b0\u5728\u7684\u5fae\u8c03\u54f2\u5b66\u662f\uff1a<\/p>\n

      \n
    • \u80fd\u7528 Prompt \u89e3\u51b3\u7684\uff0c\u7edd\u4e0d\u5fae\u8c03<\/strong><\/li>\n
    • \u80fd\u7528 RAG \u89e3\u51b3\u7684\uff0c\u7edd\u4e0d\u8bad\u7ec3<\/strong><\/li>\n
    • \u5fc5\u987b\u5fae\u8c03\u65f6\uff0c\u7528 QLoRA + \u5c0f\u6570\u636e\u96c6 + \u4f4e rank<\/strong><\/li>\n
    • \u6c38\u8fdc\u4fdd\u7559\u539f\u6a21\u578b\uff0c\u4e0d\u5408\u5e76\u6743\u91cd\uff0c\u76f4\u5230\u90e8\u7f72\u524d\u6700\u540e\u4e00\u523b<\/strong><\/li>\n<\/ul>\n

      \u6211\u751a\u81f3\u5f00\u59cb\u6000\u7591\uff1a\u672a\u6765\u4f1a\u4e0d\u4f1a\u51fa\u73b0\u2018\u6a21\u578b\u5373\u670d\u52a1\u2019\u7684\u6a21\u5f0f\u2014\u2014\u4f60\u4e0a\u4f20\u4f60\u7684\u6570\u636e\uff0c\u5e73\u53f0\u7528 QLoRA \u5728\u4e91\u7aef\u5e2e\u4f60\u5fae\u8c03\uff0c\u7136\u540e\u4f60\u53ea\u4e0b\u8f7d\u4e00\u4e2a 10MB \u7684 LoRA \u6743\u91cd\uff0c\u5c31\u80fd\u8ba9 70B \u6a21\u578b\u4e3a\u4f60\u670d\u52a1\uff1f<\/p>\n

      \u90a3\u4e00\u5929\uff0c\u53ef\u80fd\u4e0d\u8fdc\u4e86\u3002<\/p>\n

      —<\/p>\n

      \u6211\u4f9d\u7136\u5728\u5b66\u4e60\u3002\u6bcf\u4e00\u6b21\u8bad\u7ec3\uff0c\u90fd\u662f\u4e00\u6b21\u4e0e\u5de8\u517d\u7684\u5bf9\u8bdd\u3002\u6211\u4e0d\u518d\u5bb3\u6015\u5b83\u5e9e\u5927\uff0c\u56e0\u4e3a\u6211\u7ec8\u4e8e\u660e\u767d\uff1a\u771f\u6b63\u7684\u529b\u91cf\uff0c\u4e0d\u5728\u4e8e\u4f53\u79ef\uff0c\u800c\u5728\u4e8e\u5982\u4f55\u8f7b\u5de7\u5730\u64ac\u52a8\u5b83\u3002<\/p>\n

      \u56fe\u7247\u6765\u6e90<\/h2>\n
        \n
      • Personal design algorithm\u2026 I\u2019ve been trying to envision how a machine learning\/neural network can aid a designer\u2014not design things for you, but help figure out what your design should be; an idea expander; a fast sketching partner\u2026 anyway, that\u2019s been mak – bjornmeansbear, by-sa, https:\/\/www.flickr.com\/photos\/64519085@N00\/40362455043<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

        \u4ece\u5168\u91cf\u5fae\u8c03\u7684\u6602\u8d35\u4ee3\u4ef7\uff0c\u5230 LoRA \u7684\u8f7b\u91cf\u9002\u914d\uff0c\u518d\u5230 QLoRA \u5728 48GB \u663e\u5b58\u4e0a\u5fae\u8c03 65B \u6a21\u578b\u2014\u2014\u6211\u4eb2\u624b\u8e29\u8fc7\u7684\u5751\u3001\u609f\u51fa\u7684\u9053\uff0c\u548c\u90a3\u4e9b\u8bba\u6587\u6ca1\u8bf4\u7684\u5de5\u7a0b\u771f\u76f8\u3002<\/p>\n","protected":false},"author":2,"featured_media":182,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[],"class_list":["post-183","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/183","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=183"}],"version-history":[{"count":0,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/183\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/182"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=183"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=183"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=183"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}