{"id":187,"date":"2026-05-13T16:47:43","date_gmt":"2026-05-13T08:47:43","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/05\/13\/qlora%e5%be%ae%e8%b0%83%e5%8e%9f%e7%90%86%e8%af%a6%e8%a7%a3%ef%bc%9a%e4%b8%8elora%e7%9a%84%e6%80%a7%e8%83%bd%e4%b8%8e%e5%86%85%e5%ad%98%e5%af%b9%e6%af%94\/"},"modified":"2026-05-17T19:52:03","modified_gmt":"2026-05-17T11:52:03","slug":"qlora%e5%be%ae%e8%b0%83%e5%8e%9f%e7%90%86%e8%af%a6%e8%a7%a3%ef%bc%9a%e4%b8%8elora%e7%9a%84%e6%80%a7%e8%83%bd%e4%b8%8e%e5%86%85%e5%ad%98%e5%af%b9%e6%af%94","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/05\/13\/qlora%e5%be%ae%e8%b0%83%e5%8e%9f%e7%90%86%e8%af%a6%e8%a7%a3%ef%bc%9a%e4%b8%8elora%e7%9a%84%e6%80%a7%e8%83%bd%e4%b8%8e%e5%86%85%e5%ad%98%e5%af%b9%e6%af%94\/","title":{"rendered":"QLoRA\u5fae\u8c03\u539f\u7406\u8be6\u89e3\uff1a\u4e0eLoRA\u7684\u6027\u80fd\u4e0e\u5185\u5b58\u5bf9\u6bd4"},"content":{"rendered":"

\u5f15\u8a00\uff1a\u4e3a\u4ec0\u4e48\u5927\u6a21\u578b\u5fae\u8c03\u9700\u8981QLoRA\uff1f<\/h2>\n

\u5728\u6df1\u5165LoRA\u5fae\u8c03\u673a\u5236\u4e4b\u524d\uff0c\u6211\u4eec\u5fc5\u987b\u76f4\u9762\u4e00\u4e2a\u73b0\u5b9e\uff1a\u5f53\u4eca\u4e3b\u6d41\u5927\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u89c4\u6a21\u5df2\u7a81\u7834\u5343\u4ebf\u7ea7\u522b\uff0c\u4f20\u7edf\u5168\u53c2\u6570\u5fae\u8c03\u9700\u8981\u6570\u5341GB\u663e\u5b58\uff0c\u8fdc\u8d85\u666e\u901a\u5f00\u53d1\u8005\u53ef\u7528\u7684\u786c\u4ef6\u6761\u4ef6\u3002\u5373\u4fbf\u4f7f\u7528NVIDIA A100\uff0880GB\uff09\u8fd9\u7c7b\u9ad8\u7aef\u5361\uff0c\u5355\u6b21\u8bad\u7ec3\u4ecd\u9700\u5728\u5206\u5e03\u5f0f\u73af\u5883\u4e0b\u62c6\u5206\u6a21\u578b\uff0c\u590d\u6742\u5ea6\u548c\u6210\u672c\u9661\u589e\u3002<\/p>\n

LoRA\u6280\u672f\u7684\u51fa\u73b0\u66fe\u88ab\u89c6\u4e3a\u7834\u89e3\u8fd9\u4e00\u56f0\u5883\u7684\u5173\u952e\u2014\u2014\u5b83\u901a\u8fc7\u5728\u6743\u91cd\u77e9\u9635\u65c1\u8def\u63d2\u5165\u4f4e\u79e9\u66f4\u65b0\u77e9\u9635\uff0c\u4ec5\u8bad\u7ec3\u5c11\u91cf\u65b0\u589e\u53c2\u6570\uff0c\u663e\u8457\u51cf\u5c11\u4e86\u53ef\u8bad\u7ec3\u53c2\u6570\u91cf\u3002\u7136\u800c\uff0cLoRA\u5e76\u672a\u89e3\u51b3\u6a21\u578b\u52a0\u8f7d\u65f6\u7684\u5185\u5b58\u74f6\u9888\uff1a\u4e3a\u4e86\u6267\u884c\u524d\u5411\u4f20\u64ad\u548c\u68af\u5ea6\u8ba1\u7b97\uff0c\u5b8c\u6574\u6a21\u578b\u4ecd\u9700\u4ee5FP16\u7cbe\u5ea6\u52a0\u8f7d\u81f3\u663e\u5b58\uff0c\u5355\u4e2a\u5343\u4ebf\u53c2\u6570\u6a21\u578b\u5728FP16\u4e0b\u4ecd\u5360\u7528\u7ea6200GB\u5185\u5b58\uff0c\u666e\u901a\u6d88\u8d39\u7ea7\u663e\u5361\uff08\u5982RTX 3090\/4090\uff09\u6839\u672c\u65e0\u6cd5\u627f\u8f7d\u3002<\/p>\n

QLoRA\u7684\u8bde\u751f\u6b63\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u201c\u52a0\u8f7d\u5373\u7206\u663e\u5b58\u201d\u7684\u6838\u5fc3\u75db\u70b9\u3002\u5b83\u5c06LoRA\u7684\u53c2\u6570\u9ad8\u6548\u6027\u4e0e4-bit\u91cf\u5316\u6280\u672f\u6df1\u5ea6\u7ed3\u5408\uff1a\u6a21\u578b\u6743\u91cd\u88ab\u91cf\u5316\u4e3a4-bit\u6574\u6570\u5b58\u50a8\uff0c\u663e\u5b58\u5360\u7528\u538b\u7f29\u81f3\u539fFP16\u76841\/4\uff1b\u540c\u65f6\uff0c\u91cf\u5316\u8bef\u5dee\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u4e2d\u7684\u9ad8\u7cbe\u5ea6\u68af\u5ea6\u8ba1\u7b97\u8fdb\u884c\u8865\u507f\uff0c\u786e\u4fdd\u5fae\u8c03\u6027\u80fd\u51e0\u4e4e\u65e0\u635f\u3002\u5b9e\u6d4b\u8868\u660e\uff0cQLoRA\u53ef\u5728\u5355\u5f2024GB\u663e\u5b58\u7684\u6d88\u8d39\u7ea7\u663e\u5361\uff08\u5982RTX 3090\uff09\u4e0a\u5b8c\u621070B\u53c2\u6570\u6a21\u578b\u7684\u5fae\u8c03\uff0c\u800c\u6027\u80fd\u4e0e\u5168\u53c2\u6570\u5fae\u8c03\u6216\u6807\u51c6LoRA\u6301\u5e73\uff0c\u751a\u81f3\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u7565\u6709\u63d0\u5347\u3002\u8fd9\u4e00\u7a81\u7834\u4f7f\u5f97\u4e2a\u4eba\u5f00\u53d1\u8005\u4e5f\u80fd\u5728\u672c\u5730\u73af\u5883\u9ad8\u6548\u5fae\u8c03\u5927\u6a21\u578b\uff0c\u5f7b\u5e95\u6539\u53d8\u4e86\u5927\u6a21\u578b\u5e94\u7528\u7684\u95e8\u69db\u3002<\/p>\n

\u8981\u7406\u89e3QLoRA\u4e3a\u4f55\u5982\u6b64\u9ad8\u6548\uff0c\u9996\u5148\u9700\u638c\u63e1LoRA\u7684\u57fa\u672c\u539f\u7406\u2014\u2014\u5b83\u5982\u4f55\u901a\u8fc7\u4f4e\u79e9\u5206\u89e3\u5b9e\u73b0\u53c2\u6570\u538b\u7f29\u800c\u4e0d\u727a\u7272\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n

LoRA\u5fae\u8c03\u6280\u672f\u57fa\u7840\uff1a\u4f4e\u79e9\u9002\u914d\u7684\u539f\u7406<\/h2>\n

\u5728\u7406\u89e3QLoRA\u5982\u4f55\u8fdb\u4e00\u6b65\u538b\u7f29\u663e\u5b58\u5360\u7528\u4e4b\u524d\uff0c\u6211\u4eec\u5fc5\u987b\u5148\u638c\u63e1LoRA\u7684\u6838\u5fc3\u673a\u5236\u2014\u2014\u4f4e\u79e9\u9002\u914d\uff08Low-Rank Adaptation\uff09\u3002LoRA\u5e76\u975e\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u5927\u6a21\u578b\uff0c\u800c\u662f\u901a\u8fc7\u5728\u539f\u59cb\u6743\u91cd\u77e9\u9635\u65c1\u63d2\u5165\u4e00\u4e2a\u53ef\u8bad\u7ec3\u7684\u4f4e\u79e9\u589e\u91cf\u77e9\u9635\uff0c\u4ee5\u6781\u5c0f\u7684\u53c2\u6570\u5f00\u9500\u5b9e\u73b0\u9ad8\u6548\u5fae\u8c03\u3002<\/p>\n

\u5177\u4f53\u800c\u8a00\uff0c\u5047\u8bbe\u539f\u59cb\u6743\u91cd\u77e9\u9635\u4e3a W \u2208 R^{d\u00d7k}<\/code>\uff0cLoRA\u4e0d\u76f4\u63a5\u4fee\u6539 W<\/code>\uff0c\u800c\u662f\u5f15\u5165\u4e24\u4e2a\u4f4e\u7ef4\u77e9\u9635 B \u2208 R^{d\u00d7r}<\/code> \u548c A \u2208 R^{r\u00d7k}<\/code>\uff0c\u5176\u4e2d r<\/code> \u4e3a\u79e9\uff08rank\uff09\uff0c\u901a\u5e38\u53d6\u503c\u4e3a 8 \u6216 16\u3002\u5fae\u8c03\u65f6\uff0c\u6a21\u578b\u7684\u524d\u5411\u4f20\u64ad\u53d8\u4e3a\uff1a<\/p>\n

output = (W + \u0394W) \u00d7 x = W \u00d7 x + BA \u00d7 x<\/code><\/p>\n

\u5176\u4e2d \u0394W = BA<\/code> \u662f\u4f4e\u79e9\u589e\u91cf\uff0c\u5176\u603b\u53c2\u6570\u91cf\u4ec5\u4e3a d\u00d7r + r\u00d7k<\/code>\uff0c\u8fdc\u5c0f\u4e8e\u539f\u59cb\u6743\u91cd\u7684 d\u00d7k<\/code>\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u4e00\u4e2a 4096\u00d74096 \u7684\u6743\u91cd\u77e9\u9635\uff0c\u82e5\u53d6 r=8<\/code>\uff0c\u65b0\u589e\u53c2\u6570\u4ec5\u4e3a 4096\u00d78 + 8\u00d74096 = 65,536<\/code>\uff0c\u4ec5\u5360\u539f\u53c2\u6570\u7684\u7ea6 0.39%\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u539f\u59cb\u6743\u91cd W<\/code> \u88ab\u5b8c\u5168\u51bb\u7ed3\uff0c\u4ec5\u4f18\u5316 A<\/code> \u548c B<\/code>\uff0c\u5927\u5e45\u964d\u4f4e\u663e\u5b58\u9700\u6c42\u4e0e\u8ba1\u7b97\u5f00\u9500\u3002<\/p>\n

\u66f4\u91cd\u8981\u7684\u662f\uff0c\u63a8\u7406\u9636\u6bb5\u65e0\u9700\u989d\u5916\u8ba1\u7b97\uff1a\u53ea\u9700\u5c06\u589e\u91cf\u5408\u5e76\u56de\u539f\u59cb\u6743\u91cd\uff0c\u5373 W_final = W + BA<\/code>\uff0c\u6574\u4e2a\u8fc7\u7a0b\u5bf9\u7528\u6237\u900f\u660e\uff0c\u4e0d\u5f15\u5165\u4efb\u4f55\u63a8\u7406\u5ef6\u8fdf\u3002\u8fd9\u4e00\u7279\u6027\u4f7fLoRA\u5728\u5de5\u4e1a\u90e8\u7f72\u4e2d\u6781\u5177\u5438\u5f15\u529b\u3002<\/p>\n

import torch\nimport torch.nn as nn\n\nclass LoRALayer(nn.Module):\n    def __init__(self, in_features, out_features, rank=8):\n        super().__init__()\n        self.weight = nn.Linear(in_features, out_features, bias=False)\n        # \u51bb\u7ed3\u539f\u59cb\u6743\u91cd\n        self.weight.weight.requires_grad_(False)\n        # \u521d\u59cb\u5316\u4f4e\u79e9\u77e9\u9635 A \u548c B\n        self.A = nn.Linear(in_features, rank, bias=False)\n        self.B = nn.Linear(rank, out_features, bias=False)\n        # B \u521d\u59cb\u5316\u4e3a\u96f6\uff0cA \u521d\u59cb\u5316\u4e3a\u9ad8\u65af\u5206\u5e03\uff08\u7a33\u5b9a\u8bad\u7ec3\uff09\n        nn.init.zeros_(self.B.weight)\n        nn.init.normal_(self.A.weight, std=1.0 \/ rank)\n\n    def forward(self, x):\n        # \u539f\u59cb\u8f93\u51fa + \u4f4e\u79e9\u589e\u91cf\n        return self.weight(x) + self.B(self.A(x))\n\n# \u4f7f\u7528\u793a\u4f8b\uff1a\u66ff\u6362Transformer\u4e2d\u7684\u7ebf\u6027\u5c42\noriginal_linear = nn.Linear(4096, 4096)\nlora_layer = LoRALayer(4096, 4096, rank=8)\n# \u8bad\u7ec3\u65f6\u4ec5\u66f4\u65b0 A \u548c B \u7684\u53c2\u6570\nprint(f\"\u603b\u53c2\u6570: {sum(p.numel() for p in lora_layer.parameters())}\")  # \u8f93\u51fa\u7ea6 65k\nprint(f\"\u53ef\u8bad\u7ec3\u53c2\u6570: {sum(p.numel() for p in lora_layer.parameters() if p.requires_grad)}\")  # \u8f93\u51fa 65k\nprint(f\"\u51bb\u7ed3\u53c2\u6570: {sum(p.numel() for p in original_linear.parameters())}\")  # \u8f93\u51fa 16.8M\n<\/code><\/pre>\n
\u6b63\u662f\u57fa\u4e8eLoRA\u8fd9\u79cd\u8f7b\u91cf\u7ea7\u3001\u65e0\u63a8\u7406\u5f00\u9500\u7684\u9002\u914d\u673a\u5236\uff0cQLoRA\u624d\u5f97\u4ee5\u8fdb\u4e00\u6b65\u5f15\u5165\u91cf\u5316\u6280\u672f\uff0c\u5728\u4fdd\u7559\u6027\u80fd\u7684\u540c\u65f6\u5c06\u663e\u5b58\u9700\u6c42\u538b\u7f29\u81f3\u6d88\u8d39\u7ea7\u663e\u5361\u53ef\u627f\u53d7\u7684\u8303\u56f4\u3002<\/p>\n
QLoRA\u7684\u6838\u5fc3\u521b\u65b0\uff1a4-bit\u91cf\u5316\u4e0e\u53cc\u91cf\u5316\u6280\u672f<\/h2>\n
\u5728\u638c\u63e1LoRA\u901a\u8fc7\u4f4e\u79e9\u77e9\u9635\u5b9e\u73b0\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u7684\u57fa\u7840\u4e0a\uff0cQLoRA\u8fdb\u4e00\u6b65\u5f15\u5165\u4e864-bit\u91cf\u5316<\/strong>\u4e0e\u53cc\u91cf\u5316<\/strong>\u6280\u672f\uff0c\u5c06\u6a21\u578b\u6743\u91cd\u7684\u5b58\u50a8\u4e0e\u8ba1\u7b97\u5f00\u9500\u538b\u7f29\u81f3\u524d\u6240\u672a\u6709\u7684\u6c34\u5e73\uff0c\u4f7f\u5f97\u5728\u6d88\u8d39\u7ea7GPU\u4e0a\u5fae\u8c0370B\u7ea7\u5927\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n
QLoRA\u5728\u6a21\u578b\u52a0\u8f7d\u9636\u6bb5\uff0c\u5c06\u539f\u672c\u4ee5FP16\uff0816\u4f4d\u6d6e\u70b9\uff09\u5b58\u50a8\u7684\u6743\u91cd\u76f4\u63a5\u91cf\u5316\u4e3a4-bit\u6574\u6570\uff0c\u91c7\u7528\u4e00\u79cd\u540d\u4e3aNF4<\/strong>\uff08NormalFloat 4-bit\uff09\u7684\u65b0\u578b\u91cf\u5316\u683c\u5f0f\u3002\u4e0e\u4f20\u7edf\u7684\u7ebf\u6027\u91cf\u5316\u4e0d\u540c\uff0cNF4\u57fa\u4e8e\u6b63\u6001\u5206\u5e03\u7684\u6743\u91cd\u7279\u6027\u8bbe\u8ba1\uff0c\u901a\u8fc7\u975e\u5747\u5300\u91cf\u5316\u533a\u95f4\uff0c\u5728\u4f4e\u6bd4\u7279\u4e0b\u4fdd\u7559\u4e86\u66f4\u4f18\u7684\u6570\u503c\u8868\u793a\u80fd\u529b\uff0c\u5c24\u5176\u5bf9\u5927\u6a21\u578b\u4e2d\u5e38\u89c1\u7684\u957f\u5c3e\u6743\u91cd\u5206\u5e03\u5177\u6709\u66f4\u5f3a\u7684\u9c81\u68d2\u6027\u3002\u8fd9\u4e00\u8fc7\u7a0b\u5c06\u6a21\u578b\u53c2\u6570\u4f53\u79ef\u4ece\u539f\u59cbFP16\u7684\u7ea6140GB\uff0870B\u6a21\u578b\uff09\u76f4\u63a5\u538b\u7f29\u81f3\u7ea635GB\uff0c\u964d\u5e45\u8d85\u8fc775%\u3002<\/p>\n
\u4e3a\u8fdb\u4e00\u6b65\u538b\u7f29\u5185\u5b58\uff0cQLoRA\u5f15\u5165\u4e86\u53cc\u91cf\u5316\uff08Double Quantization\uff09<\/strong>\u673a\u5236\u3002\u5728\u91cf\u5316\u6743\u91cd\u65f6\uff0c\u6bcf\u4e2a\u91cf\u5316\u5757\u9700\u8981\u5b58\u50a8\u4e00\u4e2a\u7f29\u653e\u56e0\u5b50\uff08scale\uff09\u548c\u504f\u79fb\u91cf\uff08offset\uff09\uff0c\u8fd9\u4e9b\u5e38\u91cf\u672c\u8eab\u4e5f\u5360\u7528\u663e\u5b58\u3002\u53cc\u91cf\u5316\u5bf9\u8fd9\u4e9b\u91cf\u5316\u5e38\u91cf\u518d\u6b21\u8fdb\u884c4-bit\u91cf\u5316\uff0c\u5f62\u6210\u201c\u91cf\u5316\u4e2d\u7684\u91cf\u5316\u201d\uff0c\u5c06\u539f\u672c\u7528\u4e8e\u5b58\u50a8\u7f29\u653e\u56e0\u5b50\u7684\u5185\u5b58\u5f00\u9500\u518d\u51cf\u5c11\u7ea650%\u3002\u8fd9\u4e00\u6280\u672f\u770b\u4f3c\u7b80\u5355\uff0c\u5374\u5728\u4e0d\u663e\u8457\u727a\u7272\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u5b9e\u73b0\u4e86\u5185\u5b58\u5360\u7528\u7684\u8fb9\u9645\u9012\u51cf\u4f18\u5316\u3002<\/p>\n
\u4e3a\u5e94\u5bf9\u91cf\u5316\u540e\u9891\u7e41\u7684\u5f20\u91cf\u52a0\u8f7d\u4e0e\u91ca\u653e\u5bfc\u81f4\u7684\u663e\u5b58\u788e\u7247\u95ee\u9898\uff0cQLoRA\u96c6\u6210\u5206\u9875\u5185\u5b58\u7ba1\u7406\uff08PagedAttention\uff09<\/strong>\u6280\u672f\u3002\u8be5\u6280\u672f\u5c06\u6a21\u578b\u6743\u91cd\u6309\u56fa\u5b9a\u5927\u5c0f\u7684\u201c\u9875\u201d\u7ec4\u7ec7\uff0c\u52a8\u6001\u6620\u5c04\u5230\u663e\u5b58\u4e2d\uff0c\u907f\u514d\u4e86\u4f20\u7edf\u8fde\u7eed\u5185\u5b58\u5206\u914d\u5728\u591a\u8f6e\u68af\u5ea6\u66f4\u65b0\u4e2d\u4ea7\u751f\u7684\u788e\u7247\u5316\u3002\u8fd9\u4e0d\u4ec5\u63d0\u5347\u4e86\u663e\u5b58\u5229\u7528\u7387\uff0c\u66f4\u663e\u8457\u589e\u5f3a\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u7684\u7a33\u5b9a\u6027\uff0c\u5c24\u5176\u5728\u957f\u5e8f\u5217\u6216\u6279\u91cf\u8bad\u7ec3\u573a\u666f\u4e0b\uff0c\u6709\u6548\u9632\u6b62\u4e86OOM\uff08\u5185\u5b58\u6ea2\u51fa\uff09\u9519\u8bef\u7684\u53d1\u751f\u3002<\/p>\n
\u901a\u8fc7NF4\u91cf\u5316\u3001\u53cc\u91cf\u5316\u4e0e\u5206\u9875\u5185\u5b58\u7ba1\u7406\u7684\u4e09\u91cd\u534f\u540c\uff0cQLoRA\u5728\u4fdd\u6301\u4e0eLoRA\u76f8\u8fd1\u5fae\u8c03\u6027\u80fd\u7684\u540c\u65f6\uff0c\u5c06\u663e\u5b58\u9700\u6c42\u4ece\u767eGB\u7ea7\u964d\u81f3\u5355\u5361\u53ef\u627f\u8f7d\u768420GB\u4ee5\u5185\uff0c\u4e3a\u5927\u6a21\u578b\u5fae\u8c03\u5f00\u8f9f\u4e86\u5168\u65b0\u7684\u7ecf\u6d4e\u6027\u8def\u5f84\u3002<\/p>\n
\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u901a\u8fc7\u5b9e\u9a8c\u6570\u636e\u5bf9\u6bd4QLoRA\u4e0eLoRA\u5728\u76f8\u540c\u4efb\u52a1\u4e0b\u7684\u6027\u80fd\u8868\u73b0\u4e0e\u8d44\u6e90\u6d88\u8017\u5dee\u5f02\uff0c\u9a8c\u8bc1\u5176\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u6709\u6548\u6027\u3002<\/p>\n
QLoRA\u4e0eLoRA\u7684\u5185\u5b58\u4e0e\u6027\u80fd\u5bf9\u6bd4<\/h2>\n
\u5728\u7406\u89e3\u4e86QLoRA\u5982\u4f55\u901a\u8fc74-bit\u91cf\u5316\u4e0e\u53cc\u91cf\u5316\u538b\u7f29\u6a21\u578b\u6743\u91cd\u540e\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u5bf9\u6bd4\u5176\u4e0e\u4f20\u7edfLoRA\u5728\u663e\u5b58\u5360\u7528\u4e0e\u63a8\u7406\u6027\u80fd\u4e0a\u7684\u5b9e\u9645\u5dee\u5f02\uff0c\u4ee5\u63ed\u793a\u5176\u5de5\u7a0b\u4ef7\u503c\u3002<\/p>\n
\u5728\u5fae\u8c037B\u89c4\u6a21\u6a21\u578b\u65f6\uff0c\u6807\u51c6LoRA\u9700\u8981\u7ea616GB\u663e\u5b58\uff0c\u4e3b\u8981\u7528\u4e8e\u5b58\u50a8FP16\u683c\u5f0f\u7684\u4e3b\u6a21\u578b\u53c2\u6570\u4e0e\u4f4e\u79e9\u9002\u914d\u5668\u3002\u800cQLoRA\u901a\u8fc7\u5c06\u4e3b\u6a21\u578b\u6743\u91cd\u91cf\u5316\u4e3a4-bit INT8\uff08\u4f7f\u7528\u53cc\u91cf\u5316\u6280\u672f\u51cf\u5c11\u91cf\u5316\u8bef\u5dee\uff09\uff0c\u4ec5\u970010GB\u663e\u5b58\u5373\u53ef\u5b8c\u6210\u8bad\u7ec3\uff0c\u5185\u5b58\u8282\u7701\u8fbe37.5%\u3002\u66f4\u60ca\u4eba\u7684\u662f\uff0c\u5728\u6a21\u578b\u52a0\u8f7d\u9636\u6bb5\uff0cQLoRA\u53ef\u5c067B\u6a21\u578b\u7684\u663e\u5b58\u5360\u7528\u4eceLoRA\u768416GB\u8fdb\u4e00\u6b65\u538b\u7f29\u81f34GB\u2014\u2014\u8fd9\u610f\u5473\u7740\u5355\u5f20\u6d88\u8d39\u7ea7GPU\uff08\u5982RTX 3090\uff09\u5373\u53ef\u5b8c\u6574\u52a0\u8f7d\u5e76\u5fae\u8c03\u539f\u672c\u9700\u591a\u5361\u5e76\u884c\u7684\u5927\u578b\u6a21\u578b\uff0c\u5f7b\u5e95\u6253\u7834\u5927\u6a21\u578b\u5fae\u8c03\u7684\u786c\u4ef6\u95e8\u69db\u3002<\/p>\n
\u5728\u6027\u80fd\u65b9\u9762\uff0cQLoRA\u4e0d\u4ec5\u672a\u56e0\u91cf\u5316\u727a\u7272\u7cbe\u5ea6\uff0c\u53cd\u800c\u5728\u591a\u4e2a\u6743\u5a01\u57fa\u51c6\u4e0a\u8868\u73b0\u4f18\u4e8eLoRA\u3002\u5728AlpacaEval\uff08\u8bc4\u4f30\u6a21\u578b\u751f\u6210\u80fd\u529b\uff09\u548cMMLU\uff08\u591a\u4efb\u52a1\u8bed\u8a00\u7406\u89e3\uff09\u6d4b\u8bd5\u4e2d\uff0cQLoRA\u7684\u5f97\u5206\u4e0e\u5168\u53c2\u6570\u5fae\u8c03\uff08Full Fine-tuning\uff09\u7684\u5dee\u8ddd\u5c0f\u4e8e1%\uff0c\u800cLoRA\u901a\u5e38\u5b58\u57282\u20134%\u7684\u6027\u80fd\u4e0b\u964d\u3002\u8fd9\u8868\u660eQLoRA\u5728\u4fdd\u7559LoRA\u53c2\u6570\u9ad8\u6548\u6027\u7684\u540c\u65f6\uff0c\u901a\u8fc7\u66f4\u7cbe\u7ec6\u7684\u91cf\u5316\u7b56\u7565\u6709\u6548\u6062\u590d\u4e86\u6a21\u578b\u8868\u8fbe\u80fd\u529b\uff0c\u907f\u514d\u4e86\u4f20\u7edf\u4f4e\u7cbe\u5ea6\u5fae\u8c03\u4e2d\u5e38\u89c1\u7684\u68af\u5ea6\u5931\u771f\u4e0e\u4fe1\u606f\u4e22\u5931\u95ee\u9898\u3002\u5176\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\uff1a\u53cc\u91cf\u5316\uff08Double Quantization\uff09\u5c06\u91cf\u5316\u5e38\u6570\u4e5f\u4ee54-bit\u5b58\u50a8\uff0c\u663e\u8457\u964d\u4f4e\u91cf\u5316\u5e26\u6765\u7684\u989d\u5916\u5f00\u9500\uff1b\u800cNF4\uff08NormalFloat 4-bit\uff09\u8fd9\u4e00\u4e13\u4e3a\u8bed\u8a00\u6a21\u578b\u8bbe\u8ba1\u7684\u91cf\u5316\u7c7b\u578b\uff0c\u6bd4\u6807\u51c6INT4\u66f4\u9002\u5e94\u6743\u91cd\u7684\u9ad8\u65af\u5206\u5e03\uff0c\u4ece\u800c\u5728\u6781\u4f4e\u6bd4\u7279\u4e0b\u4fdd\u7559\u4e86\u5173\u952e\u4fe1\u606f\u3002<\/p>\n
\u56e0\u6b64\uff0cQLoRA\u91cd\u65b0\u5b9a\u4e49\u4e86\u201c\u9ad8\u6548\u5fae\u8c03\u201d\u7684\u6807\u51c6\u2014\u2014\u5b83\u4e0d\u518d\u662f\u7b80\u5355\u5730\u51cf\u5c11\u53ef\u8bad\u7ec3\u53c2\u6570\uff0c\u800c\u662f\u901a\u8fc7\u7cfb\u7edf\u7ea7\u7684\u91cf\u5316\u5de5\u7a0b\uff0c\u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\uff0c\u5c06\u5927\u6a21\u578b\u5fae\u8c03\u7684\u8d44\u6e90\u9700\u6c42\u538b\u7f29\u81f3\u539f\u6c34\u5e73\u7684\u56db\u5206\u4e4b\u4e00\u3002<\/p>\n
\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u6df1\u5165\u5206\u6790QLoRA\u5728\u5b9e\u9645\u8bad\u7ec3\u4e2d\u7684\u4f18\u5316\u7b56\u7565\u4e0e\u6846\u67b6\u96c6\u6210\u65b9\u5f0f\u3002<\/p>\n
QLoRA\u7684\u8bad\u7ec3\u6d41\u7a0b\u4e0e\u5de5\u7a0b\u5b9e\u73b0\u7ec6\u8282<\/h2>\n
\u5728\u638c\u63e1\u4e86QLoRA\u7684\u5185\u5b58\u4f18\u52bf\u540e\uff0c\u5b9e\u9645\u8bad\u7ec3\u6d41\u7a0b\u7684\u5de5\u7a0b\u5b9e\u73b0\u6210\u4e3a\u51b3\u5b9a\u5fae\u8c03\u6210\u8d25\u7684\u5173\u952e\u3002QLoRA\u7684\u6838\u5fc3\u5728\u4e8e\uff1a\u5728\u6781\u4f4e\u663e\u5b58\u5360\u7528\u4e0b\uff0c\u4ecd\u80fd\u4fdd\u6301\u6a21\u578b\u53c2\u6570\u7684\u9ad8\u7cbe\u5ea6\u66f4\u65b0\u80fd\u529b\uff0c\u800c\u8fd9\u4f9d\u8d56\u4e8e\u7cbe\u786e\u7684\u91cf\u5316\u51bb\u7ed3\u7b56\u7565\u4e0e\u68af\u5ea6\u4f18\u5316\u673a\u5236\u3002<\/p>\n
\u9996\u5148\uff0c\u6a21\u578b\u5fc5\u987b\u901a\u8fc7 bitsandbytes<\/code> \u5e93\u52a0\u8f7d\u4e3a4-bit\u91cf\u5316\u7248\u672c\u3002\u8be5\u5e93\u652f\u6301NF4\uff08NormalFloat 4-bit\uff09\u91cf\u5316\u683c\u5f0f\uff0c\u914d\u5408\u53cc\u91cf\u5316\uff08Double Quantization\uff09\u6280\u672f\uff0c\u5c06\u6743\u91cd\u538b\u7f29\u81f3\u7ea62GB\u4ee5\u4e0b\uff0c\u540c\u65f6\u4fdd\u7559FP16\u7cbe\u5ea6\u7684\u68af\u5ea6\u66f4\u65b0\u80fd\u529b\u3002\u8bad\u7ec3\u65f6\uff0c\u4e3b\u6a21\u578b\u7684\u91cf\u5316\u6743\u91cd\u88ab\u5b8c\u5168\u51bb\u7ed3\uff0c\u4ec5LoRA\u9002\u914d\u5668\u7684\u4f4e\u79e9\u77e9\u9635\u53c2\u4e0e\u68af\u5ea6\u8ba1\u7b97\u4e0e\u53c2\u6570\u66f4\u65b0\u3002\u8fd9\u4e00\u8bbe\u8ba1\u786e\u4fdd\u4e86\u53cd\u5411\u4f20\u64ad\u8def\u5f84\u4ec5\u4f5c\u7528\u4e8e\u5c11\u91cf\u53ef\u8bad\u7ec3\u53c2\u6570\uff0c\u5927\u5e45\u964d\u4f4e\u663e\u5b58\u5cf0\u503c\u3002<\/p>\n
\u4e3a\u8fdb\u4e00\u6b65\u538b\u7f29\u663e\u5b58\uff0c\u5fc5\u987b\u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\uff08Gradient Checkpointing\uff09\uff0c\u5b83\u901a\u8fc7\u5728\u524d\u5411\u4f20\u64ad\u65f6\u4e22\u5f03\u4e2d\u95f4\u6fc0\u6d3b\u503c\u3001\u5728\u53cd\u5411\u4f20\u64ad\u65f6\u91cd\u65b0\u8ba1\u7b97\u6765\u6362\u53d6\u5185\u5b58\u8282\u7701\uff0c\u4ee3\u4ef7\u662f\u7565\u5fae\u589e\u52a0\u8ba1\u7b97\u65f6\u95f4\u3002\u7ed3\u5408\u6279\u91cf\u5927\u5c0f\u4e3a1\u7684\u8bbe\u7f6e\uff0c\u5373\u4f7f\u5728\u5355\u5f2024GB\u663e\u5b58\u7684A100\u4e0a\uff0c\u4e5f\u80fd\u7a33\u5b9a\u8bad\u7ec37B\u89c4\u6a21\u6a21\u578b\u3002<\/p>\n
\u4f18\u5316\u5668\u63a8\u8350\u4f7f\u7528AdamW\uff0c\u5176\u6743\u91cd\u8870\u51cf\u673a\u5236\u5bf9\u5c0f\u6279\u91cf\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002\u5b66\u4e60\u7387\u8bbe\u7f6e\u4e3a 2e-4<\/code> \u662f\u7ecf\u8fc7\u5927\u91cf\u5b9e\u9a8c\u9a8c\u8bc1\u7684\u5408\u7406\u8d77\u70b9\uff0c\u8fc7\u9ad8\u6613\u5bfc\u81f4LoRA\u53c2\u6570\u9707\u8361\uff0c\u8fc7\u4f4e\u5219\u6536\u655b\u7f13\u6162\u3002\u7531\u4e8e\u5355\u6279\u6b21\u4ec51\u4e2a\u6837\u672c\uff0c\u9700\u901a\u8fc7\u68af\u5ea6\u7d2f\u79ef\uff08gradient accumulation\uff09\u6a21\u62df\u5927batch\u6548\u679c\uff0c\u901a\u5e38\u8bbe\u7f6e gradient_accumulation_steps=32<\/code>\uff0c\u4ee5\u8fbe\u5230\u7b49\u6548batch_size=32\u7684\u66f4\u65b0\u7a33\u5b9a\u6027\u3002<\/p>\n
from transformers import AutoModelForCausalLM, AutoTokenizer\nfrom peft import LoraConfig, get_peft_model\nimport bitsandbytes as bnb\n\n# \u52a0\u8f7d4-bit\u91cf\u5316\u6a21\u578b\nmodel = AutoModelForCausalLM.from_pretrained(\n    \"meta-llama\/Llama-2-7b-chat-hf\",\n    load_in_4bit=True,\n    bnb_4bit_use_double_quant=True,\n    bnb_4bit_quant_type=\"nf4\",\n    bnb_4bit_compute_dtype=torch.bfloat16\n)\n\n# \u914d\u7f6eLoRA\u9002\u914d\u5668\nlora_config = LoraConfig(\n    r=8,\n    lora_alpha=16,\n    target_modules=[\"q_proj\", \"v_proj\"],\n    lora_dropout=0.05,\n    bias=\"none\",\n    task_type=\"CAUSAL_LM\"\n)\n\nmodel = get_peft_model(model, lora_config)\n\n# \u542f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9\nmodel.gradient_checkpointing_enable()\n\n# \u8bad\u7ec3\u914d\u7f6e\noptimizer = torch.optim.AdamW(model.parameters(), lr=2e-4)\ngradient_accumulation_steps = 32\n\nfor step, batch in enumerate(dataloader):\n    outputs = model(**batch)\n    loss = outputs.loss \/ gradient_accumulation_steps\n    loss.backward()\n    \n    if (step + 1) % gradient_accumulation_steps == 0:\n        optimizer.step()\n        optimizer.zero_grad()\n<\/code><\/pre>\n
\u8be5\u5b9e\u73b0\u786e\u4fdd\u4e86\u5728\u4ec5\u4f7f\u75282\u20133GB\u989d\u5916\u663e\u5b58\u7684\u60c5\u51b5\u4e0b\uff0c\u5b8c\u6210\u5bf97B\u6a21\u578b\u7684\u9ad8\u6548\u5fae\u8c03\uff0c\u800c\u65e0\u9700\u4fee\u6539\u539f\u59cb\u6743\u91cd\u7ed3\u6784\u3002<\/p>\n
\u5c3d\u7ba1QLoRA\u5728\u8d44\u6e90\u53d7\u9650\u73af\u5883\u4e0b\u8868\u73b0\u5353\u8d8a\uff0c\u5176\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4ecd\u53d7\u5236\u4e8e\u7279\u5b9a\u786c\u4ef6\u4e0e\u6570\u636e\u5206\u5e03\u6761\u4ef6\uff0c\u4e0b\u4e00\u8282\u5c06\u6df1\u5165\u5206\u6790\u5176\u5728\u771f\u5b9e\u4efb\u52a1\u4e2d\u7684\u9002\u7528\u8fb9\u754c\u4e0e\u6f5c\u5728\u9677\u9631\u3002<\/p>\n
\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e0e\u9650\u5236\u6761\u4ef6<\/h2>\n
\u5728\u7406\u89e3\u4e86QLoRA\u7684\u8bad\u7ec3\u6d41\u7a0b\u540e\uff0c\u660e\u786e\u5176\u9002\u7528\u8fb9\u754c\u81f3\u5173\u91cd\u8981\u3002QLoRA\u6700\u663e\u8457\u7684\u4ef7\u503c\u5728\u4e8e\u8ba9\u6d88\u8d39\u7ea7GPU\uff08\u5982RTX 3090\u62164090\uff09\u80fd\u591f\u9ad8\u6548\u5fae\u8c037B\u81f313B\u53c2\u6570\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u65e0\u9700\u591a\u5361\u534f\u540c\u6216\u6602\u8d35\u7684A100\u96c6\u7fa4\u3002\u5728\u5355\u536132GB\u663e\u5b58\u73af\u5883\u4e0b\uff0c\u7528\u6237\u53ef\u5b8c\u6574\u52a0\u8f7d\u5e76\u8bad\u7ec3\u5982Llama-2-13B\u8fd9\u6837\u7684\u6a21\u578b\uff0c\u8bad\u7ec3\u901f\u5ea6\u4e0e\u5168\u53c2\u6570\u5fae\u8c03\u76f8\u6bd4\u4ec5\u4e0b\u964d10%\u201315%\uff0c\u800c\u663e\u5b58\u5360\u7528\u964d\u4f4e\u8d85\u8fc780%\u3002\u8fd9\u4e00\u7279\u6027\u6781\u5927\u964d\u4f4e\u4e86\u4e2d\u5c0f\u4f01\u4e1a\u4e0e\u7814\u7a76\u8005\u53c2\u4e0e\u5927\u6a21\u578b\u5b9a\u5236\u7684\u95e8\u69db\u3002<\/p>\n
\u7136\u800c\uff0cQLoRA\u5e76\u975e\u4e07\u80fd\u89e3\u6cd5\u3002\u5bf9\u4e8e70B\u4ee5\u4e0a\u7684\u8d85\u5927\u89c4\u6a21\u6a21\u578b\uff0c\u5373\u4fbf\u91c7\u75284-bit\u91cf\u5316\uff0c\u5355\u5361\u663e\u5b58\u4ecd\u96be\u4ee5\u627f\u8f7d\u5b8c\u6574\u6743\u91cd\u4e0e\u4f18\u5316\u5668\u72b6\u6001\uff0c\u68af\u5ea6\u7d2f\u79ef\u4e0e\u53cd\u5411\u4f20\u64ad\u7684\u5185\u5b58\u5f00\u9500\u4f1a\u8fc5\u901f\u6ea2\u51fa\u3002\u6b64\u65f6\uff0c\u63a8\u8350\u91c7\u7528\u6a21\u578b\u84b8\u998f\uff08Distillation\uff09\u6216\u6df7\u5408\u4e13\u5bb6\u67b6\u6784\uff08MoE\uff09\u4f5c\u4e3a\u66ff\u4ee3\u65b9\u6848\uff0c\u5c06\u5927\u6a21\u578b\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230\u66f4\u5c0f\u3001\u53ef\u5fae\u8c03\u7684\u5b50\u6a21\u578b\u4e2d\uff0c\u800c\u975e\u76f4\u63a5\u5bf9\u5de8\u578b\u6a21\u578b\u8fdb\u884c\u7aef\u5230\u7aef\u5fae\u8c03\u3002<\/p>\n
\u6b64\u5916\uff0c\u91cf\u5316\u8fc7\u7a0b\u4e0d\u53ef\u907f\u514d\u5730\u5f15\u5165\u8f7b\u5fae\u7cbe\u5ea6\u635f\u5931\uff0c\u5c24\u5176\u5728\u6743\u91cd\u5206\u5e03\u6781\u7aef\u6216\u68af\u5ea6\u53d8\u5316\u5267\u70c8\u7684\u5c42\u4e2d\u3002\u8fd9\u79cd\u635f\u5931\u867d\u5728\u901a\u7528\u5bf9\u8bdd\u4efb\u52a1\u4e2d\u5f71\u54cd\u5fae\u5f31\uff0c\u4f46\u5728\u91d1\u878d\u98ce\u63a7\u3001\u533b\u7597\u8bca\u65ad\u7b49\u5bf9\u8f93\u51fa\u7a33\u5b9a\u6027\u8981\u6c42\u6781\u9ad8\u7684\u573a\u666f\u4e2d\uff0c\u53ef\u80fd\u9020\u6210\u4e0d\u53ef\u63a5\u53d7\u7684\u504f\u5dee\u3002\u4f8b\u5982\uff0c\u5728\u4fe1\u8d37\u5ba1\u6279\u6a21\u578b\u4e2d\uff0c\u5fae\u8c03\u540e\u8f93\u51fa\u7684\u6982\u7387\u5206\u5e03\u82e5\u51fa\u73b0\u00b13%\u7684\u504f\u79fb\uff0c\u53ef\u80fd\u5bfc\u81f4\u5408\u89c4\u98ce\u9669\u3002\u56e0\u6b64\uff0c\u90e8\u7f72\u524d\u5fc5\u987b\u8fdb\u884c\u4e25\u683c\u7684\u6821\u51c6\u6d4b\u8bd5\u4e0e\u8bef\u5dee\u5206\u6790\uff0c\u4f18\u5148\u9009\u62e9\u9ad8\u7f6e\u4fe1\u5ea6\u8f93\u51fa\u533a\u95f4\uff0c\u6216\u7ed3\u5408\u540e\u5904\u7406\u6821\u6b63\u673a\u5236\u3002<\/p>\n
\u7efc\u4e0a\uff0cQLoRA\u662f\u5f53\u524d\u6700\u5177\u5b9e\u7528\u4ef7\u503c\u7684\u4f4e\u8d44\u6e90\u5fae\u8c03\u6280\u672f\uff0c\u4f46\u5176\u6210\u529f\u4f9d\u8d56\u4e8e\u5bf9\u4efb\u52a1\u9700\u6c42\u3001\u6a21\u578b\u89c4\u6a21\u4e0e\u7cbe\u5ea6\u5bb9\u5fcd\u5ea6\u7684\u7cbe\u51c6\u6743\u8861\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u5f15\u8a00\uff1a\u4e3a\u4ec0\u4e48\u5927\u6a21\u578b\u5fae\u8c03\u9700\u8981QLoRA\uff1f \u5728\u6df1\u5165LoRA\u5fae\u8c03\u673a\u5236\u4e4b\u524d\uff0c\u6211\u4eec\u5fc5\u987b\u76f4\u9762\u4e00\u4e2a\u73b0\u5b9e\uff1a\u5f53\u4eca\u4e3b\u6d41\u5927\u8bed\u8a00\u6a21\u578b\u7684 […]<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[],"class_list":["post-187","post","type-post","status-publish","format-standard","hentry","category-mtsb"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/187","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=187"}],"version-history":[{"count":1,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/187\/revisions"}],"predecessor-version":[{"id":204,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/187\/revisions\/204"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=187"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=187"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=187"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}