{"id":2331,"date":"2026-07-26T16:26:28","date_gmt":"2026-07-26T08:26:28","guid":{"rendered":"https:\/\/blog.zjml.life\/?p=2331"},"modified":"2026-07-26T17:26:26","modified_gmt":"2026-07-26T09:26:26","slug":"%e9%9b%b6%e5%9f%ba%e7%a1%80%e5%ae%9e%e6%88%98-lora%e5%be%ae%e8%b0%83-%e5%87%a1%e4%ba%ba%e4%bf%ae%e4%bb%99%e4%bc%a0-%e6%95%99%e7%a8%8b","status":"publish","type":"post","link":"https:\/\/blog.zjml.life\/?p=2331","title":{"rendered":"\u96f6\u57fa\u7840\u5b9e\u6218 LoRA\u5fae\u8c03 \u51e1\u4eba\u4fee\u4ed9\u4f20 \u6559\u7a0b"},"content":{"rendered":"\n
<\/i>\u00a0GitHub<\/span><\/a><\/div>
BigheadFishZJML\/qwopus3.5-9b-fanren-lora<\/span><\/a><\/div>
<\/div><\/div>
<\/i>\u00a0<\/span><\/span><\/i>\u00a0<\/span><\/span><\/div><\/div>\n\n\n\n

\n \"HuggingFace\"<\/a>\n \"GitHub\"<\/a>\n \"\u535a\u5ba2\"<\/a>\n<\/p>\n\n\n\n

\u4ec0\u4e48\u662fLoRA<\/strong>\u5fae\u8c03<\/h2><\/div>\n\n\n\n

\u4e0eLoRA<\/strong>\u5fae\u8c03\u5bf9\u5e94\u7684\u662f\u5168\u91cf\u5fae\u8c03\uff0c\u5373\u5bf9\u6574\u4e2a\u6a21\u578b\u6240\u6709\u7684\u53c2\u6570\u6839\u636e\u8f93\u5165\u7684\u6587\u672c\u8fdb\u884c\u5fae\u8c03\u66f4\u65b0\u3002\u8fd9\u4e2a\u6a21\u578b\u6240\u6709\u7684\u53c2\u6570\u9700\u8981\u5de8\u91cf\u7684\u7b97\u529b\uff0c\u53ea\u6709\u5927\u578b\u8ba1\u7b97\u673a\u7fa4\u53ef\u4ee5\u5b9e\u73b0\u3002<\/p>\n\n\n\n

LoRA<\/strong>\u5fae\u8c03\u662f\u76ee\u524d\u793e\u533a\u6700\u4e3b\u6d41\u7684\u6a21\u578b\u5fae\u8c03\u65b9\u5f0f\uff0c\u901a\u8fc7\u8bad\u7ec3\u4e24\u4e2a\u4f4e\u9636\u77e9\u9635\uff0c\u53ea\u9700\u8c03\u6574\u975e\u5e38\u5c11\u7684\u53c2\u6570\u91cf\uff0c\u80fd\u591f\u5b9e\u73b0\u8fd1\u4f3c\u5168\u91cf\u5fae\u8c03\u7684\u6548\u679c\uff0c\u5b9e\u73b0\u5728\u666e\u901a\u5bb6\u7528\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u4e5f\u80fd\u8fdb\u884c\u6a21\u578b\u7684\u5fae\u8c03\u5de5\u4f5c\u3002<\/p>\n\n\n\n

\u8bbe\u5907\u51c6\u5907<\/h2><\/div>\n\n\n\n

\u6559\u7a0b\u57fa\u4e8e\u7684\u663e\u5361\u662fRTX 5070\uff0c12GB\u663e\u5b58\u3002\u7531\u4e8e\u4e0d\u540c\u578b\u53f7\u7684\u663e\u5361\u5176\u663e\u5b58\u4e0e\u652f\u6301\u7684\u91cf\u5316\u7cbe\u5ea6\u7686\u6709\u6240\u4e0d\u540c\uff0c\u56e0\u6b64\u6709\u6761\u4ef6\u7684\u8bfb\u8005\u53ef\u5728\u81ea\u5df1\u7684\u5bb6\u7528\u7ea7\u663e\u5361\u4e0a\u8fdb\u884cLoRA\u5fae\u8c03\uff0c\u4ea6\u6216\u8005\u53ef\u4ee5\u5c1d\u8bd5\u4f7f\u7528\u963f\u91cc\u4e91\u63d0\u4f9b\u7684\u514d\u8d39\u8bbe\u5907\u4e0a\u8fdb\u884c\u5fae\u8c03\u5b66\u4e60\u3002\u672c\u6559\u7a0b\u5168\u7a0b\u65e0\u5e7f\uff0c\u7b14\u8005\u4f1a\u5728\u4e0b\u9762\u7b80\u5355\u8bb2\u8ff0\u4f7f\u7528\u963f\u91cc\u4e91\u7684\u514d\u8d39\u8bbe\u5907\u53ef\u80fd\u9047\u5230\u7684\u95ee\u9898\u548c\u5904\u7406\u65b9\u5411\u3002\u5728\u6b64\u4e4b\u540e\uff0c\u672c\u6559\u7a0b\u5c06\u4e0d\u4f1a\u6d89\u53ca\u963f\u91cc\u4e91\u7684\u8bbe\u5907\u76f8\u5173\u4fe1\u606f\u3002<\/p>\n\n\n\n

\u4f7f\u7528\u963f\u91cc\u4e91\u4eba\u5de5\u667a\u80fd\u5e73\u53f0PAI<\/a>\uff08\u975e\u5e7f\uff09<\/h3><\/div>\n\n\n\n

\u963f\u91cc\u4e91\u9996\u6708\u4f1a\u8d60\u9001\u4e00\u5b9a\u7684\u514d\u8d39\u989d\u5ea6\uff0c\u5176\u514d\u8d39\u989d\u5ea6\u5141\u8bb8\u79df\u7528\u4e09\u79cd\u7279\u5b9a\u7684\u5b9e\u4f8b\uff0c\u6211\u4eec\u9700\u8981\u79df\u8d41\u7684\u662f\u4ea4\u4e92\u5f0f\u5efa\u6a21\uff08DSW\uff09\u5b9e\u4f8b\uff0c\u7b14\u8005\u5c1d\u8bd5\u8fc7\u4f7f\u7528\u56fe\u4e2d\u9009\u62e9\u7684\u5b9e\u4f8b\u2014\u2014\u89c4\u683c\u82f1\u4f1f\u8fbeA10\u663e\u5361\uff0c\u867d\u7136\u663e\u5b58\u670924 GB\uff0c\u4f46\u7531\u4e8e\u5176\u4e0d\u652f\u6301\u66f4\u4f4e\u7684\u7cbe\u5ea6\uff0c\u6700\u7ec8\u8bad\u7ec3\u6548\u679c\u53cd\u800c\u4e0d\u5982\u7b14\u8005\u7684RTX 5070\u663e\u5361\uff0c\u56e0\u6b64\u6545\u653e\u5f03\u3002<\/p>\n\n\n\n

\u53e6\u5916\u6709\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u4e00\u70b9\u3002\u5728\u963f\u91cc\u4e91\u4eba\u5de5\u667a\u80fd\u5e73\u53f0PAI\u521b\u5efa\u5b9e\u4f8b\u7684\u65f6\u5019\uff0c\u5982\u679c\u9009\u62e9\u4e86ModelScope\u5f00\u5934\u7684\u955c\u50cf\uff0c\u5176\u6240\u6709\u7684\u8fde\u63a5Hugging Face\u5e73\u53f0\u4e0b\u8f7d\u6a21\u578b\u7684\u884c\u4e3a\u90fd\u4f1a\u88ab\u62e6\u622a\uff0c\u5fc5\u987b\u5728\u963f\u91cc\u4e91\u7684ModelScope\u5e73\u53f0\u4e0b\u8f7d\u6a21\u578b\u3002\u7531\u4e8e\u8fd9\u662f\u4e24\u4e2a\u4e0d\u540c\u7684\u5e73\u53f0\uff0c\u6240\u4ee5Hugging Face\u5e73\u53f0\u4e0a\u7684\u6a21\u578bModelScope\u5e73\u53f0\u4e0d\u4e00\u5b9a\u6709\uff0c\u56e0\u6b64\u8bfb\u8005\u5f97\u63d0\u65e9\u505a\u597d\u51c6\u5907\uff0c\u8981\u4e48\u5c31\u4e0d\u4f7f\u7528ModelScope\u5f00\u5934\u7684\u955c\u50cf\uff0c\u8981\u4e48\u5c31\u5148\u5728ModelScope\u5e73\u53f0\u9009\u62e9\u597d\u9700\u8981\u5fae\u8c03\u7684\u6a21\u578b\uff0c\u4e0d\u7136\u4f1a\u9047\u5230\u65e0\u6cd5\u8fde\u63a5Hugging Face\u4e0b\u8f7d\u6a21\u578b\u7684\u7a98\u5883\u3002<\/p>\n\n\n\n

\"\"<\/div><\/figure><\/div>\n\n\n\n

\u51c6\u5907\u5de5\u4f5c<\/h2><\/div>\n\n\n\n

\u6240\u6709\u795e\u7ecf\u7f51\u7edc\u3001\u4eba\u5de5\u667a\u80fd\u90fd\u9700\u8981\u7684\u7b2c\u4e00\u6b65\u64cd\u4f5c\uff0c\u5b89\u88c5\u8f83\u65b0\u7684\u663e\u5361\u9a71\u52a8\u7248\u672c\u3002
\u7b14\u8005\u4f7f\u7528\u7684Python\u7248\u672c\u662f3.12.13\u3002\u7531\u4e8e\u73af\u5883\u5b89\u88c5\u4e0d\u662f\u672c\u6559\u7a0b\u7684\u91cd\u70b9\uff0c\u56e0\u6b64\u7528\u6237\u53ef\u4ee5\u5c1d\u8bd5\u4f7f\u7528\u4e0b\u65b9\u7684\u65b9\u5f0f\u5feb\u901f\u914d\u7f6e\u73af\u5883\uff0c\u4f46\u7531\u4e8e\u6bcf\u4e00\u4e2a\u4eba\u6240\u4f7f\u7528\u7684\u663e\u5361\u548c\u5bf9\u5e94\u7684\u7248\u672c\u4e0d\u540c\uff0c\u82e5\u51fa\u73b0\u610f\u5916\u60c5\u51b5\uff0c\u5efa\u8bae\u4f7f\u7528DeepSeek\u7b49\u5927\u6a21\u578b\u8f85\u52a9\u73af\u5883\u5b89\u88c5\u3002<\/p>\n\n\n\n

1. \u514b\u9686\u4ed3\u5e93\n\ngit clone https:\/\/github.com\/BigheadFishZJML\/qwopus3.5-9b-fanren-lora\ncd qwopus3.5-9b-fanren-lora\n\n2. \u521b\u5efa\u865a\u62df\u73af\u5883\n\n# \u65b9\u5f0f A\uff1aconda\nconda create -n fanren python=3.12 -y\nconda activate fanren\n\n# \u65b9\u5f0f B\uff1avenv\npython -m venv venv\nvenv\\Scripts\\activate       # Windows\nsource venv\/bin\/activate    # Linux \/ macOS\n\n3. \u5b89\u88c5\u4f9d\u8d56\uff08\u6ce8\u610f\u987a\u5e8f\uff01\uff09\n\nunsloth \u4f1a\u5f3a\u5236\u62c9\u53d6 CPU \u7248 PyTorch\uff0c\u5fc5\u987b\u5148\u88c5 CUDA \u7248\u518d\u88c5 unsloth\uff1a\n\n# 1. \u5148\u88c5 CUDA \u7248 PyTorch\npip install torch==2.13.0+cu132 --index-url https:\/\/download.pytorch.org\/whl\/cu132 --force-reinstall\npip install torchvision --index-url https:\/\/download.pytorch.org\/whl\/cu132\n\n# 2. \u518d\u88c5 bitsandbytes\npip install bitsandbytes\n\n# 3. \u6700\u540e\u88c5 unsloth\uff0c\u7981\u6b62\u81ea\u52a8\u62c9\u4f9d\u8d56\npip install unsloth==2026.7.1 --no-deps\npip install unsloth-zoo==2026.7.1 --no-deps\n\n# 4. \u5176\u4f59\u4f9d\u8d56\u6b63\u5e38\u5b89\u88c5\npip install transformers trl datasets peft<\/code><\/pre>\n\n\n\n
\u7b2c\u4e00\u6b65\u2014\u2014\u521b\u5efa\u8bad\u7ec3\u96c6<\/h2><\/div>\n\n\n\n
\u9996\u5148\uff0c\u6211\u4eec\u9700\u8981\u51c6\u5907\u8bad\u7ec3\u4f7f\u7528\u7684\u8bad\u7ec3\u96c6\uff0c\u8fd9\u90e8\u5206\u7684\u4ee3\u7801\u5728split_novel.py\u4e2d\u5b9e\u73b0\u3002
\u8fd9\u4e00\u6b65\u9700\u8981\u8bfb\u8005\u81ea\u884c\u5728\u7f51\u4e0a\u51c6\u5907\u4e00\u4e2a\u51e1\u4eba\u4fee\u4ed9\u4f20\u7684txt\u6587\u672c\uff0c\u5efa\u8bae\u547d\u540d\u4e3a \u51e1\u4eba\u4fee\u4ed9\u4f20.txt
\u9879\u76ee\u4f7f\u7528 Unsloth + TRL SFTTrainer \u8fdb\u884c LoRA \u5fae\u8c03\uff0c\u56e0\u6b64\u8bad\u7ec3\u96c6\u652f\u6301\u4ee5\u4e0b\u4e09\u79cd\u8f93\u5165\u7ed3\u6784\uff1a<\/p>\n\n\n\n
\u683c\u5f0f 1\uff1ainstruction-input-output<\/p>\n\n\n\n
{\n  \"instruction\": \"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5...\",\n  \"input\": \"\u97e9\u7acb\u7ad9\u5728\u4e03\u7384\u95e8\u7684\u6f14\u6b66\u573a\u4e0a\uff0c\u5468\u56f4\u7ad9\u6ee1\u4e86\u524d\u6765\u53c2\u52a0\u5165\u95e8\u6d4b\u8bd5\u7684\u5c11\u5e74...\",\n  \"output\": \"\u8fd9\u662f\u4ed6\u7b2c\u4e00\u6b21\u9762\u5bf9\u8fd9\u4e48\u591a\u4eba\uff0c\u800c\u4e14\u8fd9\u4e9b\u4eba\u90fd\u662f\u548c\u4ed6\u5dee\u4e0d\u591a\u5927\u7684\u540c\u9f84\u4eba...\"\n}<\/code><\/pre>\n\n\n\n
\u683c\u5f0f 2\uff1aquestion-answer<\/p>\n\n\n\n
{\n  \"question\": \"\u4ec0\u4e48\u662f\u7b51\u57fa\uff1f\",\n  \"answer\": \"\u7b51\u57fa\u662f\u4fee\u4ed9\u8005\u5c06\u4f53\u5185\u771f\u6c14\u51dd\u7ed3\u4e3a\u6db2\u6001\u771f\u5143\u7684\u8fc7\u7a0b...\"\n}<\/code><\/pre>\n\n\n\n
\u683c\u5f0f 3\uff1a\u7eaf\u6587\u672c<\/p>\n\n\n\n
{\n  \"text\": \"\u97e9\u7acb\u76d8\u819d\u5750\u5728\u6d1e\u5e9c\u4e4b\u4e2d\uff0c\u53cc\u624b\u6390\u8bc0...\"\n}<\/code><\/pre>\n\n\n\n
\u7b14\u8005\u8fd9\u91cc\u91c7\u7528\u4e86\u7b2c\u4e00\u79cd\u683c\u5f0f\uff0c\u4f46\u4f5c\u4e3a\u4e00\u4e2a\u7eed\u5199\u6a21\u578b\uff0c\u5176\u5b9e\u4f7f\u7528\u7b2c\u4e8c\u79cd\u66f4\u597d\uff0c\u4f46\u662f\u7531\u4e8e\u9879\u76ee\u4e00\u5f00\u59cb\u4f7f\u7528\u4e86\u7b2c\u4e00\u79cd\u683c\u5f0f\uff0c\u90a3\u4e48\u5c31\u5148\u6682\u65f6\u4e0d\u52a8\u4e86\u3002\u4e3a\u4ec0\u4e48\u7406\u8bba\u4e0a\u4f7f\u7528\u7b2c\u4e8c\u79cd\u683c\u5f0f\u66f4\u597d\u5462\uff1f\u56e0\u4e3a\u8fd9\u4e2a\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u4e4b\u540e\uff0c\u6211\u4eec\u8bbe\u8ba1\u5b83\u5c31\u53ea\u662f\u4e00\u4e2a\u7528\u4e8e\u6269\u5199\u7684\u6a21\u578b\u3002\u90a3\u5b83\u5c31\u4e0d\u9700\u8981\u518d\u6709\u4e00\u4e2a\u6c38\u8fdc\u90fd\u662f\u91cd\u590d\u7684\u5185\u5bb9\u7684instruction\u4e86\u3002\u8fd9\u4e2ainstruction\u4f1a\u5360\u7528\u4e00\u70b9\u4e0a\u4e0b\u6587\uff0c\u5bf9\u4e8e\u8fd9\u79cd\u4e0d\u5927\u7684\u6a21\u578b\uff0c\u80fd\u8282\u7ea6\u4e00\u70b9\u4e0a\u4e0b\u6587\u5f53\u7136\u662f\u8282\u7ea6\u4e00\u70b9\u66f4\u597d\u3002<\/p>\n\n\n\n
\u9996\u5148\u5bfc\u5165\u9700\u8981\u7528\u5230\u7684\u5305\uff1a<\/p>\n\n\n\n
import json\nimport argparse\nimport os\nfrom typing import List, Dict<\/code><\/pre>\n\n\n\n
\u63a5\u4e0b\u6765\u6211\u4eec\u5199\u4e00\u4e2a\u65b9\u6cd5\u7528\u4e8e\u52a0\u8f7d\u6587\u672c\uff0c\u56e0\u4e3a\u6211\u7684\u6587\u672c\u662fUTF-8\uff08\u4f60\u7684\u4e5f\u5f88\u53ef\u80fd\u662f\u7684\uff09\uff0c\u56e0\u6b64\u6307\u5b9a\u6587\u672c\u7f16\u7801\u4e3aUTF-8
\u8fd9\u4e2a\u65b9\u6cd5\u8f93\u5165\u4e00\u4e2a\u6587\u672c\u6587\u4ef6\u7684\u8def\u5f84\uff0c\u8fd4\u56de\u6587\u672c\u6587\u4ef6\u7684\u5b57\u7b26\u4e32\u5f62\u5f0f:<\/p>\n\n\n\n
def load_text(file_path: str) -> str:\n    \"\"\"\u8bfb\u53d6\u6587\u672c\u6587\u4ef6\uff0c\u652f\u6301UTF-8\u7f16\u7801\"\"\"\n    with open(file_path, 'r', encoding='utf-8') as f:\n        return f.read()<\/code><\/pre>\n\n\n\n
\u6709\u4e86\u6587\u672c\uff0c\u6211\u4eec\u8fd8\u9700\u8981\u5199\u4e00\u4e2a\u65b9\u6cd5\u5c06\u6587\u672c\u5207\u5272\u6210\u8bad\u7ec3\u6837\u672c\uff0c
\u8bad\u7ec3\u6837\u672c\u5e94\u8be5\u662f\u7a97\u53e3\u6ed1\u52a8\u7684\uff0c\u8fd9\u6837\u80fd\u6700\u5927\u9650\u5ea6\u5730\u5229\u7528\u6211\u4eec\u7684\u6587\u672c\uff0c\u521b\u5efa\u51fa\u66f4\u591a\u7684\u8bad\u7ec3\u6837\u672c\uff0c\u8ba9\u6a21\u578b\u8fdb\u884c\u5b66\u4e60\uff0c\u9996\u5148\u8ba9\u6211\u4eec\u5148\u7406\u89e3\u4e00\u4e0b\uff0c\u4ec0\u4e48\u662f\u7a97\u53e3\u6ed1\u52a8\uff1f<\/p>\n\n\n\n
\u4f5c\u4e3a\u8ba1\u7b97\u673a\u9886\u57df\u4e00\u4e2a\u6bd4\u8f83\u5e38\u89c1\u7684\u672f\u8bed\uff0c\u7a97\u53e3\u6ed1\u52a8\u5176\u5b9e\u5f88\u597d\u7406\u89e3\uff0c\u5bf9\u4e8e\u683c\u5f0f 1\u7684\u6837\u672c\uff0c\u6211\u4eec\u6559\u6a21\u578b\u8f93\u5165instruction\u548cinput\u7684\u5185\u5bb9\uff0c\u8981\u6c42\u8fd4\u56deoutput\u7684\u5185\u5bb9\uff0c\u800c\u7a97\u53e3\u6ed1\u52a8\u5c31\u662f\u672c\u6837\u672c\u7684input\u7684\u5185\u5bb9\u548coutput\u7684\u5185\u5bb9\uff0c\u76f8\u5bf9\u4e0b\u4e00\u4e2a\u6837\u672c\u7684input\u7684\u5185\u5bb9\u548coutput\u7684\u5185\u5bb9\uff0c\u5411\u540e\u6ed1\u52a8\u591a\u5c11\u5b57\u7b26\uff08\u6b65\u957f\uff09\uff0c\u4ee5\u4e0b\u6211\u4e3e\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
\u6837\u672c 1\uff1a<\/p>\n\n\n\n
{\n  \"instruction\": \"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5...\",\n  \"input\": \"ABCDEFG\",\n  \"output\": \"HIJKLMN\"\n}<\/code><\/pre>\n\n\n\n
\u6837\u672c 2:<\/p>\n\n\n\n
{\n  \"instruction\": \"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5...\",\n  \"input\": \"DEFGHIJ\",\n  \"output\": \"KLMN......\"\n}<\/code><\/pre>\n\n\n\n
\u90a3\u4e48\u8fd9\u91cc\u7684\u6ed1\u52a8\u7a97\u53e3\u6b65\u957f\u5c31\u662f3\uff0c\u56e0\u4e3a\u672c\u6837\u672c\u7684input\u7684\u5185\u5bb9\u548coutput\u7684\u5185\u5bb9\uff0c\u76f8\u5bf9\u4e0b\u4e00\u4e2a\u6837\u672c\u7684input\u7684\u5185\u5bb9\u548coutput\u7684\u5185\u5bb9\uff0c\u5411\u540e\u6ed1\u52a8\u4e863\u4e2a\u5b57\u7b26\u3002\u5728\u672c\u6559\u7a0b\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u8bbe\u7f6e\u6ed1\u52a8\u6b65\u957f\u7b49\u4e8einput\u7684\u957f\u5ea6\uff0c\u8fd9\u6837\u5c31\u53ef\u4ee5\u6b63\u597d\u505a\u5230\u4e0d\u91cd\u590d\uff0c\u5373\u4ee5\u4e0b\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
\u6837\u672c 1\uff1a<\/p>\n\n\n\n
{\n  \"instruction\": \"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5...\",\n  \"input\": \"ABCDEFG\",\n  \"output\": \"HIJKLMN\"\n}<\/code><\/pre>\n\n\n\n
\u6837\u672c 2:<\/p>\n\n\n\n
{\n  \"instruction\": \"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5...\",\n  \"input\": \"HIJKLMN\",\n  \"output\": \"OPQ RST......\"\n}<\/code><\/pre>\n\n\n\n
\u7406\u89e3\u4e86\u6837\u672c\u7684\u7ec4\u7ec7\u5f62\u5f0f\uff0c\u6211\u4eec\u6765\u5b9e\u73b0\u8fd9\u4e2a\u65b9\u6cd5\uff1a<\/p>\n\n\n\n
def split_text(text: str,\n               input_len: int,\n               output_len: int,\n               step: int,\n               instruction: str) -> List[Dict[str, str]]:\n    \"\"\"\n    \u5c06\u6587\u672c\u5207\u5272\u6210\u8bad\u7ec3\u6837\u672c\u3002\n    - input_len: \u6bcf\u4e2a\u6837\u672c\u7684\u8f93\u5165\uff08\u524d\u6587\uff09\u957f\u5ea6\uff08\u5b57\u7b26\u6570\uff09\n    - output_len: \u6bcf\u4e2a\u6837\u672c\u7684\u8f93\u51fa\uff08\u7eed\u5199\uff09\u957f\u5ea6\uff08\u5b57\u7b26\u6570\uff09\n    - step: \u7a97\u53e3\u6ed1\u52a8\u6b65\u957f\uff08\u5b57\u7b26\u6570\uff09\n    - instruction: \u56fa\u5b9a\u7684\u6307\u4ee4\u6587\u672c\n    \"\"\"\n    total_len = len(text)\n    samples = []\n    start = 0\n\n    while start + input_len + output_len <= total_len:\n        input_text = text[start: start + input_len]\n        output_text = text[start + input_len: start + input_len + output_len]\n\n        samples.append({\n            \"instruction\": instruction,\n            \"input\": input_text,\n            \"output\": output_text\n        })\n        start += step\n\n    return samples<\/code><\/pre>\n\n\n\n
\u8fd9\u4e2a\u65b9\u6cd5\u6700\u540e\u4f1a\u8fd4\u56de\u4e00\u4e2a\u6837\u672c\u96c6\uff0c\u975e\u5e38\u4e0d\u9519\uff01<\/p>\n\n\n\n
\u73b0\u5728\u7ec8\u4e8e\u51c6\u5907\u597d\u9700\u8981\u7528\u5230\u7684\u65b9\u6cd5\u4e86\uff0c\u73b0\u5728\u518d\u8ba9\u6211\u4eec\u5199\u4e00\u4e2amain()\u65b9\u6cd5\uff0c\u7528\u4e8e\u8c03\u7528\u6211\u4eec\u7684\u8fd9\u4e9b\u5de5\u5177\uff1a<\/p>\n\n\n\n
def main():\n    parser = argparse.ArgumentParser(description=\"\u5c06\u5c0f\u8bf4TXT\u5207\u5272\u4e3aJSON\u8bad\u7ec3\u6570\u636e\")\n\n    # \u5141\u8bb8\u901a\u8fc7\u547d\u4ee4\u8c03\u7528\u811a\u672c\u7684\u65f6\u5019\uff0c\u624b\u52a8\u8bbe\u7f6e\u6587\u672c\u957f\u5ea6\u3001\u6587\u4ef6\u8def\u5f84\u7b49\u8bbe\u7f6e\u503c\n    parser.add_argument(\"--input\", \"-i\", default=\"\u51e1\u4eba\u4fee\u4ed9\u4f20.txt\",\n                        help=\"\u8f93\u5165TXT\u6587\u4ef6\u8def\u5f84\uff08\u9ed8\u8ba4: \u51e1\u4eba\u4fee\u4ed9\u4f20.txt\uff09\")\n    parser.add_argument(\"--output\", \"-o\", default=\"fan_ren.json\",\n                        help=\"\u8f93\u51faJSON\u6587\u4ef6\u8def\u5f84\uff08\u9ed8\u8ba4: fan_ren.json\uff09\")\n    parser.add_argument(\"--input-len\", type=int, default=380,\n                        help=\"\u8f93\u5165\u6587\u672c\u957f\u5ea6\uff0c\u5b57\u7b26\u6570\uff08\u9ed8\u8ba4: 380\uff09\")\n    parser.add_argument(\"--output-len\", type=int, default=380,\n                        help=\"\u8f93\u51fa\u6587\u672c\u957f\u5ea6\uff0c\u5b57\u7b26\u6570\uff08\u9ed8\u8ba4: 380\uff09\")\n    parser.add_argument(\"--step\", type=int, default=None,\n                        help=\"\u6ed1\u52a8\u6b65\u957f\uff0c\u9ed8\u8ba4\u7b49\u4e8e input-len\uff08\u4e0d\u91cd\u53e0\uff09\")\n    parser.add_argument(\"--instruction\", type=str,\n                        default=\"\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5\uff0c\u8981\u6c42\u6587\u98ce\u4e00\u81f4\uff0c\u903b\u8f91\u8fde\u8d2f\u3002\",\n                        help=\"\u56fa\u5b9a\u7684\u6307\u4ee4\u6587\u672c\")\n\n    args = parser.parse_args()\n\n    # \u5982\u679c\u672a\u6307\u5b9astep\uff0c\u5219\u9ed8\u8ba4\u4e0d\u91cd\u53e0\uff0c\u5373\u6b65\u957f\u7b49\u4e8einput_len\n    if args.step is None:\n        args.step = args.input_len\n\n    print(f\"\u8bfb\u53d6\u6587\u4ef6: {args.input}\")\n    text = load_text(args.input)\n    total_len = len(text)\n    print(f\"\u6587\u672c\u603b\u957f\u5ea6: {total_len} \u5b57\u7b26\")\n\n    if total_len < args.input_len + args.output_len:\n        print(\"\u9519\u8bef\uff1a\u6587\u672c\u957f\u5ea6\u4e0d\u8db3\u4e00\u4e2a\u6837\u672c\uff0c\u8bf7\u51cf\u5c0finput-len\u6216output-len\")\n        return\n\n    samples = split_text(text, args.input_len, args.output_len, args.step, args.instruction)\n    print(f\"\u751f\u6210\u6837\u672c\u6570: {len(samples)}\")\n\n    # \u5199\u5165JSON\u6587\u4ef6\uff0c\u4fdd\u8bc1\u4e2d\u6587\u4e0d\u8f6c\u4e49\n    with open(args.output, 'w', encoding='utf-8') as f:\n        json.dump(samples, f, ensure_ascii=False, indent=2)\n\n    print(f\"\u6570\u636e\u5df2\u4fdd\u5b58\u81f3: {args.output}\")\n    if len(samples) > 0:\n        print(\"\u793a\u4f8b\u6837\u672c\uff08\u524d100\u4e2a\u5b57\u7b26\u7684input\uff09:\")\n        print(samples[0]['input'][:100] + \"...\")\n        print(\"\u793a\u4f8b\u6837\u672c\uff08\u524d100\u4e2a\u5b57\u7b26\u7684output\uff09:\")\n        print(samples[0]['output'][:100] + \"...\")<\/code><\/pre>\n\n\n\n
\u5728\u8fd9\u91cc\u6211\u8bbe\u7f6e\u7684\u8f93\u5165\u6587\u672c\u7684\u957f\u5ea6\u548c\u8f93\u51fa\u6587\u672c\u7684\u957f\u5ea6\uff0c\u90fd\u662f380\u4e2a\u5b57\u7b26\uff0c\u8fd9\u662f\u7b14\u8005\u5728\u8fd9\u91cc\u6298\u4e2d\u91c7\u7528\u7684\u4e00\u4e2a\u957f\u5ea6\u3002\u5bf9\u4e8e\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u5176\u8f93\u5165\u7684\u6587\u672c\u8d8a\u957f\uff0c\u5b83\u5728\u8bad\u7ec3\u4e2d\u6240\u5360\u7528\u7684\u663e\u5b58\u5c31\u8981\u8d8a\u591a\uff0c\u800c\u5f80\u5f80\u5e76\u4e0d\u662f\u6837\u672c\u8d8a\u957f\u8d8a\u597d\uff0c\u5982\u679c\u6837\u672c\u5f88\u957f\uff0c\u5176\u6837\u672c\u6570\u91cf\u5c31\u8981\u88ab\u8feb\u51cf\u5c11\uff0c\u4e0d\u4e00\u5b9a\u6709\u52a9\u4e8e\u6837\u672c\u7684\u5fae\u8c03\u3002\u5728\u8fd9\u4e2a\u9879\u76ee\u4e2d\uff0c\u6211\u4eec\u5e0c\u671b\u5b9e\u73b0\u7684\u662f\u4e00\u4e2a\u51e1\u4eba\u4fee\u4ed9\u4f20\u98ce\u683c\u7684\u7eed\u5199\u6a21\u578b\uff0c\u7531\u4e8e\u5176\u57fa\u5ea7\u6a21\u578b\u4e0d\u662f\u4e00\u4e2a\u5927\u6a21\u578b\uff0c\u6211\u4eec\u4e0d\u4f1a\u5bf9\u5176\u957f\u8f93\u51fa\u6709\u8fc7\u9ad8\u7684\u671f\u671b\u3002\u56e0\u6b64\uff0c\u4e0e\u5176\u901a\u8fc7\u957f\u6587\u672c\u8fdb\u884c\u8bad\u7ec3\uff0c\u4e0d\u5982\u901a\u8fc7\u8f83\u77ed\u7684\u6837\u672c\u8fdb\u884c\u5fae\u8c03\uff0c\u4f7f\u5176\u8f93\u51fa\u5728300~500\u5b57\u7b26\u4e4b\u95f4\u7684\u8d28\u91cf\u66f4\u9ad8\u3002<\/p>\n\n\n\n
\u6700\u540e\uff0c\u8ba9\u6211\u4eec\u6765\u786e\u8ba4\u5165\u53e3\u65b9\u6cd5\uff0c\u521b\u5efa\u8bad\u7ec3\u96c6\u7684\u4ee3\u7801\u5c31\u5199\u5b8c\u4e86\uff1a<\/p>\n\n\n\n
if __name__ == \"__main__\":\n    main()<\/code><\/pre>\n\n\n\n
\u73b0\u5728\u8ba9\u6211\u4eec\u6765\u5c1d\u8bd5\u8fd0\u884c\u8fd9\u4e2a\u811a\u672c\u3002\u522b\u5fd8\u4e86\u51c6\u5907\u597d \u51e1\u4eba\u4fee\u4ed9\u4f20.txt\uff1a<\/p>\n\n\n\n
(lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>python split_novel.py\n\u8bfb\u53d6\u6587\u4ef6: \u51e1\u4eba\u4fee\u4ed9\u4f20.txt\n\u6587\u672c\u603b\u957f\u5ea6: 7614083 \u5b57\u7b26\n\u751f\u6210\u6837\u672c\u6570: 20036\n\u6570\u636e\u5df2\u4fdd\u5b58\u81f3: fan_ren.json\n\u793a\u4f8b\u6837\u672c\uff08\u524d100\u4e2a\u5b57\u7b26\u7684input\uff09:\n\u5185\u5bb9\u7b80\u4ecb\n\u3000\u3000\u4e00\u4e2a\u666e\u901a\u7684\u5c71\u6751\u7a77\u5c0f\u5b50\uff0c\u97e9\u7acb\uff0c\u5728\u81ea\u5df1\u53d4\u53d4\u7684\u5f15\u8350\u4e0b\uff0c\u8d70\u51fa\u5c71\u91ce\uff0c\u8de8\u5165\u4e86\u4e00\u4e2a\u6c5f\u6e56\u5c0f\u95e8\u6d3e\uff0c\u6210\u4e3a\u4e86\u4e00\u540d\u957f\u8001\u7684\u8bb0\u540d\u5f1f\u5b50\uff0c\u4e00\u4e2a\u9634\u8c0b\uff0c\u8ba9\u4ed6\u61f5\u61f5\u61c2\u61c2\u7684\u5f00\u59cb\u4fee\u4ed9\uff0c\u4ece\u6b64\uff0c\u98d8\u6e3a\u65e0\u8e2a\u7684\u4ed9\u8def\u51fa\u73b0\u5728\u4e86\u4ed6\u7684\u811a\u4e0b\uff0c\u7834\u89e3\u95e8\u6d3e...\n\u793a\u4f8b\u6837\u672c\uff08\u524d100\u4e2a\u5b57\u7b26\u7684output\uff09:\n\u53cc\u773c\uff0c\u76f4\u76f4\u671b\u7740\u8305\u8349\u548c\u70c2\u6ce5\u7cca\u6210\u7684\u9ed1\u5c4b\u9876\uff0c\u8eab\u4e0a\u76d6\u7740\u7684\u65e7\u68c9\u88ab\uff0c\u5df2\u5448\u6df1\u9ec4\u8272\uff0c\u770b\u4e0d\u51fa\u539f\u6765\u7684\u672c\u6765\u9762\u76ee\uff0c\u8fd8\u82e5\u6709\u82e5\u65e0\u7684\u6563\u53d1\u7740\u6de1\u6de1\u7684\u9709\u5473\u3002\n\u3000\u3000\u5728\u4ed6\u8eab\u8fb9\u7d27\u6328\u7740\u7684\u53e6\u4e00\u4eba\uff0c\u662f\u4e8c\u54e5\u97e9\u94f8\uff0c\u9163\u7761\u7684\u5341\u5206\u9999\u751c\uff0c\u4ece\u4ed6\u8eab\u4e0a\u4e0d\u65f6\u4f20\u6765\u8f7b\u91cd\u4e0d...<\/code><\/pre>\n\n\n\n
\u6210\u529f\u4e86\uff01\u53ef\u4ee5\u770b\u5230\u6211\u4eec\u751f\u6210\u7684\u6837\u672c\u96c6\u5df2\u7ecf\u653e\u5728\u4e86fan_ren.json\uff0c\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u7528\u8fd9\u4e2a\u6837\u672c\u96c6\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u4e86\u3002<\/p>\n\n\n\n
\u7b2c\u4e8c\u6b65\u2014\u2014\u7f16\u5199\u5fae\u8c03\u811a\u672c<\/h2><\/div>\n\n\n\n
\u7ec8\u4e8e\u51c6\u5907\u597d\u4e86\u6837\u672c\u96c6\uff0c\u73b0\u5728\u8fdb\u5165\u6700\u91cd\u8981\u7684\u4e00\u6b65\uff0c\u7f16\u5199LoRA\u5fae\u8c03\u7684\u811a\u672c\u3002<\/p>\n\n\n\n
\u5728\u672c\u9879\u76ee\u4e2d\uff0c\u7b14\u8005\u9009\u62e9\u7684\u6a21\u578b\u662fzjml\/Qwen3.5-9B-Text-Only-abliterated\u3002\u8fd9\u4e2a\u6a21\u578b\u662f\u7b14\u8005\u5728Qwen3.5-9B-abliterated\u6a21\u578b\u7684\u57fa\u7840\u4e0a\u5265\u79bb\u89c6\u89c9\u5854\uff0c\u53cd\u6b63\u90fd\u662f\u672c\u5730\u8fd0\u884c\u7684\u5c0f\u6a21\u578b\uff0c\u81ea\u7136\u9009\u62e9abliterated\u7248\u672c\uff0c\u6211\u53ef\u4e0d\u60f3\u6a21\u578b\u5199\u7740\u5199\u7740\u7f62\u5de5\u3002\u81f3\u4e8e\u4e3a\u4ec0\u4e48\u8981\u4f7f\u7528\u5265\u79bb\u89c6\u89c9\u5854\u7684\u6a21\u578b\uff0c\u7531\u4e8e\u672c\u9879\u76ee\u5e0c\u671b\u5b9e\u73b0\u7684\u662f\u51e1\u4eba\u4fee\u4ed9\u4f20\u98ce\u683c\u7684\u6587\u672c\u7eed\u5199\u6a21\u578b\uff0c\u81ea\u7136\u65e0\u9700\u56fe\u7247\u7684\u8bc6\u522b\u80fd\u529b\u3002\u867d\u7136\u7b14\u8005\u4f30\u8ba19B\u7684\u6a21\u578b\u53ca\u56fe\u7247\u8bc6\u522b\u80fd\u529b\uff0c\u53ef\u80fd\u8f83\u96be\u4ee5\u80dc\u4efb\u65e5\u5e38\u5de5\u4f5c\u3002\u53e6\u5916\uff0c\u7b14\u8005\u4eb2\u6d4b\u4f7f\u7528\u5265\u79bb\u89c6\u89c9\u5854\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u5728\u8bad\u7ec3\u4e2d\u8282\u7ea6\u5927\u69820.9 GB\u7684\u663e\u5b58\uff0c\u5e76\u4e14\u80fd\u7565\u5fae\u63d0\u9ad8\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n\n\n\n
\u672c\u9879\u76ee\u91c7\u7528Unsloth \u6838\u5fc3\u5e93\uff0cUnsloth \u662f\u201c\u652f\u6301 LoRA \u7684\uff0c\u4e14\u628a\u5b83\u4f18\u5316\u5230\u6781\u81f4\u7684\u5de5\u5177\u201d\u3002\u5176\u91cd\u5199\u4e86\u6a21\u578b\u5e95\u5c42\u8ba1\u7b97\u65f6\u7684\u4e00\u4e9b\u5173\u952e\u8fd0\u7b97\u903b\u8f91\u3002PyTorch \u9ed8\u8ba4\u7684\u8ba1\u7b97\u65b9\u5f0f\u6bd4\u8f83\u201c\u89c4\u77e9\u201d\u4e14\u8017\u663e\u5b58\uff0c\u800c Unsloth \u56e2\u961f\u4e3a Llama\u3001Qwen\u3001Mistral \u7b49\u4e3b\u6d41\u6a21\u578b\u624b\u5199\u4e86\u6781\u5176\u7cbe\u7b80\u7684\u201c\u81ea\u5b9a\u4e49\u5185\u6838\uff08Custom Kernels\uff09\u201c\uff0c\u4f7f\u5728\u6d88\u8d39\u7ea7\u663e\u5361\u4e0a\u8bad\u7ec3\u5927\u6a21\u578b\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n\n\n\n
1. \u5bfc\u5165\u5305<\/h3><\/div>\n\n\n\n
\u9996\u5148\u5bfc\u5165\u9700\u8981\u4f7f\u7528\u7684\u5305\uff0c\u5728\u8fd9\u4e2a\u811a\u672c\u7684\u7f16\u5199\u4e2d\uff0c\u7b14\u8005\u9047\u5230\u4e86\u7f16\u7801\u7684\u95ee\u9898\u3002\u6839\u636e\u7b14\u8005\u7684\u6d4b\u8bd5\uff0c\u5fc5\u987b\u8981\u628a\u5f3a\u5236UTF-8\u7f16\u7801\u5199\u5728\u6240\u6709\u7b2c\u4e09\u65b9\u5305\u4e4b\u524d\uff0c\u5426\u5219\u5b9e\u9645\u8fd0\u884c\u7684\u65f6\u5019\uff0c\u4f1a\u56e0\u4e3a\u7f16\u7801\u62a5\u9519\uff1a<\/p>\n\n\n\n
import sys\n\n# Windows \u4e0a\u5f3a\u5236 UTF-8 \u6a21\u5f0f\uff0c\u5fc5\u987b\u5728 import \u4efb\u4f55\u7b2c\u4e09\u65b9\u5e93\u4e4b\u524d\nif sys.flags.utf8_mode == 0:\n    import os\n    os.environ[\"PYTHONUTF8\"] = \"1\"\n    os.execv(sys.executable, [sys.executable] + sys.argv)\nelse:\n    import os\nimport time\nimport json\nimport logging\nimport traceback\nfrom datetime import datetime\nfrom pathlib import Path<\/code><\/pre>\n\n\n\n
2. \u914d\u7f6e\u65e5\u5fd7<\/h3><\/div>\n\n\n\n
\u672c\u6587\u5728\u5bfc\u5165unsloth\u7b49\u91cd\u91cf\u7ea7\u5e93\u524d\uff0c\u5148\u5b9e\u73b0\u4e86\u65e5\u5fd7\u90e8\u5206\u7684\u5185\u5bb9\uff0c\u5728\u5bfc\u5165torch\u5305\u540e\u5148\u8f93\u51fa\u65e5\u5fd7\uff0c\u53ef\u4ee5\u65b9\u4fbf\u7684\u770b\u5230\u4f7f\u7528\u8005\u5b89\u88c5\u7684torch\u7248\u672c\u662f\u5426\u6b63\u786e\uff0c\u5982\u679c\u8f93\u51fa\u663e\u793atorch\u7248\u672c\u662fCPU\u7248\u672c\uff0c\u5219\u53ef\u5feb\u901f\u5b9a\u4f4d\u5230torch\u7248\u672c\u7684\u95ee\u9898\uff0c\u5426\u5219\uff0c\u4e4b\u540eunsloth\u5305\u5bfc\u5165\u65f6\uff0c\u7531\u4e8e\u9700\u8981\u5e26GPU\u7684torch\u7248\u672c\u800c\u62a5\u9519\uff0c\u5219\u9700\u8981\u5728\u4e00\u5927\u6bb5\u62a5\u9519\u4e2d\u5bfb\u627e\u539f\u56e0\u4e86\u3002<\/p>\n\n\n\n
\u73b0\u5728\uff0c\u8ba9\u6211\u4eec\u6765\u5b9e\u73b0\u65e5\u5fd7\u90e8\u5206\uff1a<\/p>\n\n\n\n
# ===================== \u65e5\u5fd7\u7cfb\u7edf =====================\nLOG_DIR = Path(\".\/logs\")\nLOG_DIR.mkdir(exist_ok=True)\nLOG_FILE = LOG_DIR \/ f\"train_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log\"\n\nlogging.basicConfig(\n    level=logging.INFO,\n    format=\"%(asctime)s | %(levelname)-7s | %(message)s\",\n    datefmt=\"%H:%M:%S\",\n    handlers=[\n        logging.FileHandler(LOG_FILE, encoding=\"utf-8\"),\n        logging.StreamHandler(sys.stdout),\n    ],\n)\nlog = logging.getLogger(__name__)\n\ndef log_section(title: str):\n    msg = f\"{'='*60}\\n  {title}\\n{'='*60}\"\n    log.info(msg)\n\ndef log_system_info():\n    import torch\n    log.info(f\"Python:    {sys.version}\")\n    log.info(f\"PyTorch:   {torch.__version__}  |  CUDA: {torch.version.cuda}\")\n    gpu_name = torch.cuda.get_device_name(0)\n    log.info(f\"GPU:       {gpu_name}\")\n    mem = torch.cuda.get_device_properties(0).total_memory \/ 1024**3\n    log.info(f\"VRAM \u603b\u91cf: {mem:.1f} GB\")\n    log.info(f\"\u65e5\u5fd7\u6587\u4ef6:  {LOG_FILE.resolve()}\")\n    for k in [\"PYTHONUTF8\", \"CUDA_VISIBLE_DEVICES\"]:\n        log.info(f\"  env {k} = {os.environ.get(k, '(\u672a\u8bbe\u7f6e)')}\")\n    return gpu_name, mem<\/code><\/pre>\n\n\n\n
\u63a5\u4e0b\u6765\u6211\u4eec\u8c03\u7528log_system_info()\u65b9\u6cd5\uff0c\u770b\u770b\u6211\u4eec\u7684\u73af\u5883\u662f\u5426\u6b63\u5e38\uff1a<\/p>\n\n\n\n
# ===== \u5148\u8bb0\u5f55\u73af\u5883\uff0c\u518d\u5bfc\u5165\u91cd\u91cf\u7ea7\u5e93 =====\nimport torch\ngpu_name, vram_gb = log_system_info()<\/code><\/pre>\n\n\n\n
\u8ba9\u6211\u4eec\u6765\u8fd0\u884c\u4e00\u4e0b\uff1a<\/p>\n\n\n\n
(lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>11:45:42 | INFO    | Python:    3.12.13 | packaged by Anaconda, Inc. | (main, Mar 19 2026, 20:16:45) [MSC v.1942 64 bit (AMD64)]\n11:45:42 | INFO    | PyTorch:   2.13.0+cu132  |  CUDA: 13.2\n11:45:42 | INFO    | GPU:       NVIDIA GeForce RTX 5070\n11:45:42 | INFO    | VRAM \u603b\u91cf: 11.9 GB\n11:45:42 | INFO    | \u65e5\u5fd7\u6587\u4ef6:  D:\\Data\\Python\\qwopus3.5-9b-fanren-lora\\logs\\train_20260724_114535.log\n11:45:42 | INFO    |   env PYTHONUTF8 = 1\n11:45:42 | INFO    |   env CUDA_VISIBLE_DEVICES = (\u672a\u8bbe\u7f6e)<\/code><\/pre>\n\n\n\n
\u53ef\u4ee5\u770b\u5230\uff0c\u7b14\u8005\u7684PyTorch\u7248\u672c\u662f\u5e26GPU\u7684\u7248\u672c\uff0c\u8bfb\u8005\u670b\u53cb\u4eec\u5982\u679c\u53d1\u73b0\u8fd9\u91cc\u6ca1\u6709+cuxxx\uff0c\u5219\u9700\u8981\u6ce8\u610f\u4e86\uff0c\u5f88\u53ef\u80fd\u662f\u5b89\u88c5\u7684PyTorch\u7248\u672c\u4e0d\u5bf9\u5bfc\u81f4\u7684\uff0c\u7531\u4e8eunsloth\u662f\u9700\u8981\u5e26GPU\u7684\u7248\u672cPyTorch\u7684\uff0c\u5982\u679cPyTorch\u7248\u672c\u4e0d\u5bf9\uff0c\u4e4b\u540e\u5bfc\u5165\u5305\u4f1a\u62a5\u9519\u3002<\/p>\n\n\n\n
\u73b0\u5728\u6211\u4eec\u628a\u5269\u4f59\u7684\u5305\u5bfc\u5165\u9879\u76ee\uff1a<\/p>\n\n\n\n
import datasets as hf_datasets\nfrom unsloth import FastLanguageModel\nfrom trl import SFTTrainer, SFTConfig\nfrom transformers import TrainerCallback, EarlyStoppingCallback<\/code><\/pre>\n\n\n\n
3. \u914d\u7f6e\u5168\u5c40\u914d\u7f6e<\/h3><\/div>\n\n\n\n
\u5173\u4e8e\u7b2c\u4e09\u65b9\u5305\u7684\u5bfc\u5165\u7ec8\u4e8e\u505a\u5b8c\u4e86\uff0c\u73b0\u5728\u6211\u4eec\u6765\u914d\u7f6e\u4e00\u4e9b\u5168\u5c40\u53c2\u6570\u3002<\/p>\n\n\n\n
# ===================== \u5168\u5c40\u914d\u7f6e =====================\n# \u4f18\u5148\u4f7f\u7528\u672c\u5730\u6a21\u578b\uff0c\u4e0d\u5b58\u5728\u5219\u4ece HuggingFace \u81ea\u52a8\u4e0b\u8f7d\nLOCAL_MODEL_DIR  = str(Path(__file__).resolve().parent \/ \"model\/Qwen3.5-9B-Text-Only-abliterated\")\nHF_MODEL_ID      = \"zjml\/Qwen3.5-9B-Text-Only-abliterated\"\nMODEL_ID         = LOCAL_MODEL_DIR if Path(LOCAL_MODEL_DIR).exists() else HF_MODEL_ID\n\nLORA_SAVE_DIR    = \".\/lora\/qwen3.5-9b-text-lora-checkpoint\"\nJSON_DATA_PATH   = \".\/fan_ren.json\"<\/code><\/pre>\n\n\n\n
\u63a5\u4e0b\u6765\u9700\u8981\u914d\u7f6e\u7684\u9009\u53d6\u53c2\u6570\u5c31\u4e0e\u6a21\u578b\u7684\u5fae\u8c03\u606f\u606f\u76f8\u5173\u4e86\uff0c\u6211\u5f85\u4f1a\u4f1a\u9010\u4e00\u505a\u51fa\u89e3\u91ca\uff1a<\/p>\n\n\n\n
# ---- \u6839\u636e GPU \u81ea\u52a8\u9009\u62e9\u914d\u7f6e ----\nIS_V100 = \"V100\" in gpu_name or \"Tesla V100\" in gpu_name\n\nif IS_V100:\n    if vram_gb >= 30:\n        # V100 32GB: \u957f\u4e0a\u4e0b\u6587 + \u5927 batch\n        MAX_SEQ_LENGTH            = 2048\n        TRAIN_BATCH_SIZE          = 2\n        GRADIENT_ACCUMULATION_STEPS = 4   # effective batch = 8\n    else:\n        # V100 16GB: \u4e2d\u7b49\u4e0a\u4e0b\u6587\n        MAX_SEQ_LENGTH            = 1536\n        TRAIN_BATCH_SIZE          = 1\n        GRADIENT_ACCUMULATION_STEPS = 4   # effective batch = 4\nelse:\n    # RTX 5070 \/ \u5176\u4ed6 12GB \u5361\n    MAX_SEQ_LENGTH            = 1280\n    TRAIN_BATCH_SIZE          = 1\n    GRADIENT_ACCUMULATION_STEPS = 4\n\nLEARNING_RATE             = 2e-4\nNUM_TRAIN_EPOCHS          = 3\nSAVE_STEPS                = 500\nVAL_SPLIT_RATIO           = 0.05  # 5% \u505a\u9a8c\u8bc1\u96c6\nEARLY_STOPPING_PATIENCE   = 3     # eval_loss \u8fde\u7eed 3 \u6b21\u4e0d\u964d\u5219\u505c\nLORA_DROPOUT              = 0.05  # \u8f7b\u5fae\u6b63\u5219\u5316\u9632\u8fc7\u62df\u5408<\/code><\/pre>\n\n\n\n
MAX_SEQ_LENGTH\uff1a\u5373\u6a21\u578b\u7684\u8f93\u5165\u5b57\u7b26\u6570\uff0c\u8bb0\u5f97\u6211\u4eec\u521a\u521a\u5728\u5f62\u6210\u6837\u672c\u96c6\u7684\u65f6\u5019\u8bbe\u7f6e\u7684\u5b57\u7b26\u6570\u5417\uff1f\u5f53\u65f6\u6211\u4eec\u8bbe\u7f6e\u7684\u8f93\u5165\u6587\u672c\u957f\u5ea6\u548c\u8f93\u51fa\u6587\u672c\u957f\u5ea6\u90fd\u662f380\uff0c\u518d\u52a0\u4e0a\u4e00\u6bb5introduction\uff0c\u5728\u4f7f\u75285070\u663e\u5361\u65f6\uff0c\u6211\u4eec\u5c06\u6a21\u578b\u7684\u8f93\u5165\u5b57\u7b26\u6570\u8bbe\u7f6e\u4e3a1280\uff0c\u57fa\u672c\u6b63\u597d\u80fd\u591f\u8986\u76d6\u6211\u4eec\u6240\u6709\u7684\u6837\u672c\u7684\u603b\u5b57\u7b26\u957f\u5ea6\u3002\u5982\u679c\u8fd9\u4e2a\u503c\u5c0f\u4e8e\u6837\u672c\u7684\u603b\u5b57\u7b26\u6570\uff0c\u6837\u672c\u5c06\u4f1a\u88ab\u622a\u65ad\u3002\u5f53\u7136\u4e5f\u662f\u53ef\u4ee5\u8fdb\u884c\u8bad\u7ec3\u7684\uff0c\u4e0d\u8fc7\u8fd9\u5c31\u9020\u6210\u4e86\u6837\u672c\u8d44\u6e90\u7684\u6d6a\u8d39\uff0c\u6240\u4ee5\u6837\u672c\u96c6\u5728\u8bbe\u7f6e\u8f93\u5165\u6587\u672c\u957f\u5ea6\u548c\u8f93\u51fa\u6587\u672c\u957f\u5ea6\u65f6\uff0c\u4e0e\u6211\u4eec\u73b0\u5728\u8bbe\u7f6e\u7684\u6a21\u578b\u7684\u8f93\u5165\u5b57\u7b26\u6570\u662f\u606f\u606f\u76f8\u5173\u7684\u3002<\/p>\n\n\n\n
TRAIN_BATCH_SIZE\uff1a\u5373\u8bad\u7ec3\u6279\u6b21\u5927\u5c0f\uff0c\u6a21\u578b\u4e00\u6b21\u53ea\u201c\u6d88\u5316\u201d1\u6761\u6570\u636e\uff0c\u5c31\u66f4\u65b0\u4e00\u6b21\u601d\u8003\u65b9\u5411\u3002\u5982\u679c\u663e\u5b58\u5145\u8db3\uff0c\u8bbe\u5927\u4e00\u70b9\uff08\u59824\u62168\uff09\u8bad\u7ec3\u4f1a\u66f4\u7a33\u5b9a\uff0c\u4f46\u663e\u5b58\u5360\u7528\u4f1a\u7ffb\u500d\u3002\u8fd9\u4e00\u6761\u4e0e\u4e0b\u4e00\u6761GRADIENT_ACCUMULATION_STEPS\u914d\u7f6e\u4e5f\u662f\u606f\u606f\u76f8\u5173\u7684\u3002<\/p>\n\n\n\n
GRADIENT_ACCUMULATION_STEPS\uff1a\u5373\u68af\u5ea6\u7d2f\u79ef\u6b65\u6570\uff0c\u867d\u7136\u4e00\u6b21\u53ea\u770b1\u6761\u6570\u636e\uff08\u8bad\u7ec3\u6279\u6b21\u5927\u5c0f=1\uff09\uff0c\u4f46\u6211\u628a4\u6b21\u201c\u770b\u6570\u636e\u201d\u7b97\u51fa\u6765\u7684\u68af\u5ea6\u6512\u5728\u4e00\u8d77\uff0c\u51d1\u591f\u4e86\u518d\u7edf\u4e00\u66f4\u65b0\u4e00\u6b21\u6a21\u578b\u3002\u8fd9\u6837\u5b9e\u9645\u7b49\u6548\u7684\u6279\u6b21\u5927\u5c0f = 1 \u00d7 4 = 4\uff0c\u76f8\u5f53\u4e8e\u7528\u201c\u591a\u82b1\u70b9\u65f6\u95f4\u201d\u6362\u53d6\u4e86\u201c\u5927\u6279\u6b21\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u201d\uff0c\u540c\u65f6\u8282\u7701\u4e86\u663e\u5b58\u3002<\/p>\n\n\n\n
LEARNING_RATE\uff1a\u5373\u5b66\u4e60\u7387\uff0c\u6a21\u578b\u6bcf\u6b21\u66f4\u65b0\u53c2\u6570\u65f6\u8fc8\u7684\u201c\u6b65\u5b50\u201d\u6709\u591a\u5927\u30022e-4\uff08\u5373 0.0002\uff09\u662fLoRA\u5fae\u8c03\u7684\u6807\u51c6\u9ec4\u91d1\u503c\u3002<\/p>\n\n\n\n
NUM_TRAIN_EPOCHS\uff1a\u5373\u8bad\u7ec3\u8f6e\u6570\uff0c\u76f8\u5f53\u4e8e\u6570\u636e\u96c6\u53ef\u4ee5\u91cd\u590d\u4f7f\u7528\u591a\u5c11\u904d\uff0c\u8fd9\u91cc\u8bbe\u7f6e\u4e3a\u4e863\u904d\uff0c\u76f8\u5f53\u4e8e\u5bf9\u4e8e\u6a21\u578b\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\u5927\u5c0f\u589e\u52a0\u5230\u4e863\u500d\uff0c\u4f46\u662f\u4e4b\u540e\u6211\u4eec\u8fd8\u4f1a\u8bbe\u7f6e\u65e9\u505c\uff0c\u4e0e\u65e9\u505c\u914d\u5408\u8d77\u6765\u4f7f\u7528\uff0c\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\u5927\u5c0f\u662f\u7ef0\u7ef0\u6709\u4f59\u7684\u3002\u5bf9\u4e8e\u5c0f\u6570\u636e\u96c6\uff08\u51e0\u5343\u6761\uff09\uff0c3-5\u8f6e\u5408\u9002\uff1b\u5bf9\u4e8e\u5927\u6570\u636e\u96c6\uff08\u51e0\u4e07\u6761\uff09\uff0c1-2\u8f6e\u5c31\u591f\u4e86\u3002\u8f6e\u6570\u592a\u591a\u4f1a\u628a\u6a21\u578b\u201c\u80cc\u4e0b\u6765\u201d\u5bfc\u81f4\u8fc7\u62df\u5408\u3002<\/p>\n\n\n\n
SAVE_STEPS\uff1a\u5373\u4fdd\u5b58\u68c0\u67e5\u70b9\uff0c\u6bcf\u66f4\u65b0500\u6b65\uff0c\u81ea\u52a8\u628a\u5f53\u524d\u7684\u6a21\u578b\u6743\u91cd\u5b58\u4e00\u4efd\u5230\u786c\u76d8\u4e0a\uff0c\u9632\u6b62\u8bad\u7ec3\u4e2d\u65ad\u5bfc\u81f4\u767d\u8dd1\u3002<\/p>\n\n\n\n
VAL_SPLIT_RATIO\uff1a\u5373\u9a8c\u8bc1\u96c6\u6bd4\u4f8b\uff0c\u628a\u6570\u636e\u96c6\u76845%\u62bd\u51fa\u6765\u4f5c\u4e3a\u6d4b\u8bd5\uff0c\u8fd9\u90e8\u5206\u7684\u7b54\u6848\u6a21\u578b\u6c38\u8fdc\u770b\u4e0d\u89c1\uff0c\u53ea\u7528\u6765\u68c0\u6d4b\u5b83\u5b66\u6ca1\u5b66\u61c2\u3002<\/p>\n\n\n\n
EARLY_STOPPING_PATIENCE\uff1a\u5373\u65e9\u505c\u8010\u5fc3\u503c\uff0c\u5982\u679c\u201c\u9a8c\u8bc1\u96c6Loss\u201d\u8fde\u7eed3\u6b21\u90fd\u6ca1\u6709\u53d8\u597d\uff08\u6ca1\u4e0b\u964d\uff09\uff0c\u76f4\u63a5\u6309\u4e0b\u6682\u505c\u952e\uff0c\u7ec8\u6b62\u8bad\u7ec3\uff0c\u9632\u6b62\u6a21\u578b\u8fc7\u62df\u5408\u3002<\/p>\n\n\n\n
LORA_DROPOUT\uff1a\u5373LoRA\u968f\u673a\u5931\u6d3b\u7387\uff0c\u5728LoRA\u63d2\u4ef6\u91cc\uff0c\u968f\u673a\u201c\u6254\u6389\u201d5%\u7684\u795e\u7ecf\u5143\u8fde\u63a5\uff0c\u4e0d\u8ba9\u5b83\u4eec\u53c2\u4e0e\u672c\u6b21\u66f4\u65b0\uff0c\u9632\u6b62\u6a21\u578b\u201c\u6b7b\u8bb0\u786c\u80cc\u201d\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n\n\n\n
4. \u5f00\u59cb\u4e3b\u6d41\u7a0b<\/h3><\/div>\n\n\n\n
\u7ec8\u4e8e\u914d\u7f6e\u597d\u4e86\uff01\u63a5\u4e0b\u6765\u6211\u4eec\u6765\u5199\u5fae\u8c03\u8bad\u7ec3\u7684\u4e3b\u6d41\u7a0b\u3002
\u6211\u4eec\u5148\u5199\u4e00\u4e2a\u65b9\u6cd5\u7528\u4e8e\u6e05\u7406GPU\u663e\u5b58\u788e\u7247\uff0c\u5426\u5219\u6839\u636e\u7b14\u8005\u7684\u5b9e\u6d4b\uff0c\u5176\u663e\u5b58\u5360\u7528\u4f1a\u8d8a\u6765\u8d8a\u9ad8\u3002\u4e00\u65e6\u89e6\u53d1\u4e86\u81ea\u52a8\u5378\u8f7d\uff0cPyTorch\u4f1a\u5c06\u663e\u5b58\u653e\u4e0d\u4e0b\u7684\u6570\u636e\u79fb\u52a8\u5230\u5185\u5b58\u4e2d\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u901f\u5ea6\u7684\u4e25\u91cd\u53d8\u6162\uff0c\u5bfc\u81f4\u539f\u672c\u53ea\u9700\u657010\u5c0f\u65f6\u7684\u8bad\u7ec3\u9700\u8981\u5ef6\u957f\u5230\u6570\u5929\u4e43\u81f3\u6570\u5468\uff0c\u8fd9\u662f\u65e0\u6cd5\u8ba9\u4eba\u63a5\u53d7\u7684\u3002<\/p>\n\n\n\n
# ===================== \u4e3b\u6d41\u7a0b =====================\n\nclass ClearCacheCallback(TrainerCallback):\n    \"\"\"\u6bcf\u6b21 checkpoint \u4fdd\u5b58\u540e\u6e05\u7406 GPU \u663e\u5b58\u788e\u7247\"\"\"\n    def on_save(self, args, state, control, **kwargs):\n        torch.cuda.empty_cache()\n        log.info(\"   \ud83e\uddf9 GPU cache \u5df2\u6e05\u7406\")<\/code><\/pre>\n\n\n\n
\u63a5\u4e0b\u6765\u6211\u4eec\u8981\u5f00\u59cb\u5199main()\u65b9\u6cd5\uff0c\u5728\u5199main()\u65b9\u6cd5\u4e4b\u524d\uff0c\u6211\u4eec\u9700\u8981\u7406\u6e05\u695a\u4e09\u4e2a\u4e1c\u897f\uff0c\u5206\u522b\u662f4-bit\u91cf\u5316\u3001fp16\u548cbf16\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u6d41\u7a0b\u6765\u7406\u89e3\u8fd9\u4e09\u8005\u7684\u5173\u7cfb\uff1a<\/p>\n\n\n\n
\u5f53\u4f60\u5728RTX 5070\u4e0a\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801\uff0c\u6267\u884c\u4e00\u6b21\u63a8\u7406\u65f6\uff0c\u6d41\u7a0b\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n\n\n\n
    \n
  1. \u52a0\u8f7d\uff08\u5b58\u653e\uff09<\/strong>\uff1a\u6a21\u578b\u6743\u91cd\u4ee5 4-bit<\/strong> \u683c\u5f0f\u52a0\u8f7d\u5230\u663e\u5b58\u3002\uff08\u5360\u7528\u7ea6\u539f\u672c1\/8\u7684\u7a7a\u95f4\uff0c\u6240\u4ee5\u4f60\u7684RTX 5070\u624d\u80fd\u8dd1\u52a8\u5927\u6a21\u578b\uff09<\/li>\n\n\n\n
  2. \u8ba1\u7b97\u524d\uff08\u89e3\u538b\uff09<\/strong>\uff1aGPU\u8ba1\u7b97\u5355\u5143\u8981\u7b97\u4e00\u4e2a\u77e9\u9635\u4e58\u6cd5\u3002\u5b83\u4ece\u663e\u5b58\u4e2d\u53d6\u51fa4-bit\u7684\u6743\u91cd\u3002<\/li>\n\n\n\n
  3. \u8ba1\u7b97\u4e2d\uff08\u8fd0\u7b97\uff09<\/strong>\uff1aGPU\u5185\u90e8\u4e34\u65f6<\/strong>\u628a\u8fd9\u4e9b4-bit\u6570\u5b57\u6620\u5c04\uff08\u53cd\u91cf\u5316\uff09\u6210 BF16<\/strong> \u683c\u5f0f\uff0c\u7136\u540e\u4ee5 BF16<\/strong> \u7684\u7cbe\u5ea6\u8fdb\u884c\u4e58\u6cd5\u52a0\u6cd5\u8fd0\u7b97\u3002<\/li>\n\n\n\n
  4. \u8ba1\u7b97\u540e\uff08\u5b58\u56de\uff09<\/strong>\uff1a\u7b97\u5b8c\u5f97\u5230\u7684\u7ed3\u679c\uff08\u4e5f\u662f BF16<\/strong> \uff09\uff0cGPU\u4f1a\u518d\u6b21\u628a\u5b83\u538b\u7f29\u56de4-bit<\/strong>\uff0c\u5b58\u56de\u663e\u5b58\u91cc\uff0c\u4ee5\u4fbf\u817e\u51fa\u7a7a\u95f4\u7ed9\u4e0b\u4e00\u5c42\u3002<\/li>\n<\/ol>\n\n\n\n
    \u56e0\u4e3a V100 \u7684\u786c\u4ef6\u7535\u8def\u7269\u7406\u4e0a\u4e0d\u652f\u6301<\/strong> BF16 \u7684\u8ba1\u7b97\u6307\u4ee4\u3002\u56e0\u6b64\u9700\u914d\u7f6e\u5176\u4f7f\u7528 FP16<\/strong> \u683c\u5f0f\u3002<\/p>\n\n\n\n
    5. \u52a0\u8f7d\u6a21\u578b<\/h3><\/div>\n\n\n\n
    \u73b0\u5728\u6211\u4eec\u5f00\u59cb\u5199main()\u65b9\u6cd5\uff0c\u6211\u4eec\u5148\u8981\u628a\u57fa\u5ea7\u6a21\u578b\u52a0\u8f7d\u597d\uff0c\u57fa\u5ea7\u6a21\u578b\u5982\u679c\u5b58\u653e\u5728\u672c\u5730\uff0c\u5219\u4f18\u5148\u4ece\u672c\u5730\u8c03\u53d6\u6a21\u578b\uff0c\u5426\u5219\u8bbe\u7f6e\u5176\u4eceHugging Face\u4e0b\u8f7d\u6a21\u578b\uff0c\u5e76\u4e14\u6211\u4eec\u9700\u8981\u6839\u636e\u4e0d\u540c\u7684\u663e\u5361\u914d\u7f6e\uff0c\u4f7f\u7528\u4e0d\u540c\u7684\u6570\u636e\u683c\u5f0f\uff1a<\/p>\n\n\n\n
        log_section(\"STEP 1: \u52a0\u8f7d\u6a21\u578b (4-bit \u91cf\u5316)\")\n\n    source = \"\u672c\u5730\" if MODEL_ID == LOCAL_MODEL_DIR else \"HuggingFace\"\n    log.info(f\"   \u6a21\u578b\u6765\u6e90: {source} ({MODEL_ID})\")\n\n    # V100 \u4e0d\u652f\u6301 bf16\uff0c\u4f7f\u7528 fp16\n    compute_dtype = torch.float16 if IS_V100 else torch.bfloat16\n    dtype_name = \"fp16\" if IS_V100 else \"bf16\"\n    log.info(f\"   \u8ba1\u7b97\u7cbe\u5ea6: {dtype_name}\uff08{'V100 \u4e0d\u652f\u6301 bf16' if IS_V100 else '\u81ea\u52a8\u9009\u62e9'}\uff09\")<\/code><\/pre>\n\n\n\n
    \u914d\u7f6e\u597d\u52a0\u8f7d\u6a21\u578b\u7684\u53c2\u6570\u4e4b\u540e\uff0c\u6211\u4eec\u5f00\u59cb\u5199\u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668\u7684\u4ee3\u7801\uff1a<\/p>\n\n\n\n
        t0 = time.time()\n    model, tokenizer = FastLanguageModel.from_pretrained(\n        model_name=MODEL_ID,\n        max_seq_length=MAX_SEQ_LENGTH,\n        dtype=compute_dtype,\n        load_in_4bit=True,\n    )\n    elapsed = time.time() - t0\n    log.info(f\"\u2705 \u6a21\u578b\u52a0\u8f7d\u6210\u529f\uff01\u8017\u65f6 {elapsed:.1f}s\")\n    log.info(f\"   GPU \u663e\u5b58\u5360\u7528: {torch.cuda.memory_allocated(0)\/1024**3:.1f} GB\")<\/code><\/pre>\n\n\n\n
    \u6211\u4eec\u73b0\u5728\u6765\u8fd0\u884c\u4e00\u4e0b\uff0c\u770b\u770b\u80fd\u4e0d\u80fd\u6b63\u5e38\u628a\u6a21\u578b\u52a0\u8f7d\u51fa\u6765\u3002\u7531\u4e8e\u7b14\u8005\u7684\u6a21\u578b\u662f\u76f4\u63a5\u4e0b\u8f7d\u5b58\u653e\u5230\u4e86\u672c\u5730\uff0c\u56e0\u6b64\u8fd9\u4e00\u6b65\u6bd4\u8f83\u5feb\uff0c\u5982\u679c\u8bfb\u8005\u6ca1\u6709\u63d0\u524d\u5c06\u6a21\u578b\u4e0b\u8f7d\u5230\u672c\u5730\uff0c\u7a0b\u5e8f\u9700\u8981\u81ea\u884c\u4eceHugging Face\u4e0a\u4e0b\u8f7d\u6a21\u578b\uff0c\u53ef\u80fd\u8f83\u6162\u3002\u5efa\u8bae\u8bfb\u8005\u5148\u81ea\u884c\u5c06\u6a21\u578b\u4e0b\u8f7d\u5230\u672c\u5730\uff0c\u9632\u6b62\u8fd9\u4e00\u6b65\u65f6\u95f4\u8fc7\u957f\u3002<\/p>\n\n\n\n
    \u4e3a\u4e86\u9875\u9762\u770b\u8d77\u6765\u6bd4\u8f83\u7b80\u6d01\uff0c\u7b14\u8005\u5220\u9664\u4e86\u4e00\u4e9b\u4e0d\u5f71\u54cd\u540e\u7eed\u6b65\u9aa4\u7684\u62a5\u9519\u3002\u8fd9\u4e00\u6b65\u7b14\u8005\u5927\u6982\u7b49\u5f85\u4e861~2\u5206\u949f\uff0c\u7531\u4e8e\u8bbe\u5907\u7684\u4e0d\u540c\uff0c\u7b49\u5f85\u65f6\u95f4\u4e5f\u53ef\u80fd\u6709\u6240\u4e0d\u540c\uff0c\u4f46\u4f30\u8ba1\u4e0d\u53ef\u80fd\u5927\u4e8e10\u5206\u949f\uff1a<\/p>\n\n\n\n
    (lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>python .\/train.py\n\n(lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>15:18:20 | INFO    | Python:    3.12.13 | packaged by Anaconda, Inc. | (main, Mar 19 2026, 20:16:45) [MSC v.1942 64 bit (AMD64)]\n15:18:20 | INFO    | PyTorch:   2.13.0+cu132  |  CUDA: 13.2\n15:18:20 | INFO    | GPU:       NVIDIA GeForce RTX 5070\n15:18:20 | INFO    | VRAM \u603b\u91cf: 11.9 GB\n15:18:20 | INFO    | \u65e5\u5fd7\u6587\u4ef6:  D:\\Data\\Python\\qwopus3.5-9b-fanren-lora\\logs\\train_20260725_151814.log\n15:18:20 | INFO    |   env PYTHONUTF8 = 1\n15:18:20 | INFO    |   env CUDA_VISIBLE_DEVICES = (\u672a\u8bbe\u7f6e)\n\ud83e\udda5 Unsloth: Will patch your computer to enable 2x faster free finetuning.\n\ud83e\udda5 Unsloth Zoo will now patch everything to make training faster!\n<string>:1: FutureWarning: torch._dynamo.config.inline_inbuilt_nn_modules is deprecated and does not do anything, inline_inbuilt_nn_modules is always True. It will be removed in a future version of PyTorch.\n15:19:19 | INFO    | ============================================================\n  \u73af\u5883\u8bca\u65ad\n============================================================\n15:19:19 | INFO    | ============================================================\n  STEP 1: \u52a0\u8f7d\u6a21\u578b (4-bit \u91cf\u5316)\n============================================================\n15:19:19 | INFO    |    \u6a21\u578b\u6765\u6e90: \u672c\u5730 (D:\\Data\\Python\\qwopus3.5-9b-fanren-lora\\model\\Qwen3.5-9B-Text-Only-abliterated)\n15:19:19 | INFO    |    \u8ba1\u7b97\u7cbe\u5ea6: bf16\uff08\u81ea\u52a8\u9009\u62e9\uff09\n15:19:20 | INFO    | HTTP Request: GET https:\/\/huggingface.co\/api\/agent-harnesses \"HTTP\/1.1 200 OK\"\n15:19:20 | INFO    | HTTP Request: GET https:\/\/huggingface.co\/api\/whoami-v2 \"HTTP\/1.1 200 OK\"\n15:19:20 | INFO    | HTTP Request: GET https:\/\/huggingface.co\/api\/whoami-v2 \"HTTP\/1.1 200 OK\"\n15:19:23 | INFO    | HTTP Request: GET https:\/\/huggingface.co\/api\/whoami-v2 \"HTTP\/1.1 200 OK\"\n==((====))==  Unsloth 2026.7.1: Fast Qwen3_5 patching. Transformers: 5.13.1.\n   \\\\   \/|    NVIDIA GeForce RTX 5070. Num GPUs = 1. Max memory: 11.94 GB. Platform: Windows.\nO^O\/ \\_\/ \\    Torch: 2.13.0+cu132. CUDA: 12.0. CUDA Toolkit: 13.2. Triton: 3.7.1\n\\        \/    Bfloat16 = TRUE. FA [Xformers = None. FA2 = False]\n \"-____-\"     Free license: http:\/\/github.com\/unslothai\/unsloth\nUnsloth: Fast downloading is enabled - ignore downloading bars which are red colored!\nThe fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https:\/\/github.com\/fla-org\/flash-linear-attention#installation and https:\/\/github.com\/Dao-AILab\/causal-conv1d\nLoading weights:   0%|\u258e                                                                | 2\/427 [00:03<11:33,  1.63s\/it]D:\\Data\\Anacondea\\envs\\lora\\Lib\\site-packages\\bitsandbytes\\backends\\cuda\\ops.py:213: FutureWarning: _check_is_size will be removed in a future PyTorch release along with guard_size_oblivious.     Use _check(i >= 0) instead.\n  torch._check_is_size(blocksize)\nLoading weights: 100%|\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| 427\/427 [00:14<00:00, 29.34it\/s]\nThe tokenizer you are loading from 'D:\\Data\\Python\\qwopus3.5-9b-fanren-lora\\model\\Qwen3.5-9B-Text-Only-abliterated' with an incorrect regex pattern: https:\/\/huggingface.co\/mistralai\/Mistral-Small-3.1-24B-Instruct-2503\/discussions\/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue.\n15:19:43 | INFO    | \u2705 \u6a21\u578b\u52a0\u8f7d\u6210\u529f\uff01\u8017\u65f6 23.7s\n15:19:43 | INFO    |    GPU \u663e\u5b58\u5360\u7528: 7.2 GB<\/code><\/pre>\n\n\n\n
    6. \u914d\u7f6eLoRA\u53c2\u6570<\/h3><\/div>\n\n\n\n
    \u63a5\u4e0b\u6765\u6211\u4eec\u8981\u914d\u7f6eLoRA\u5fae\u8c03\u7684\u53c2\u6570\uff0c\u8fd9\u4e00\u6b65\u6bd4\u8f83\u67af\u71e5\uff0c\u4e0d\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u6211\u914d\u7f6e\u597d\u7684\u3002<\/p>\n\n\n\n
    r=16\uff1aLoRA \u4f4e\u79e9\u77e9\u9635\u7684\u79e9<\/strong>\uff0c\u4e3a\u4ec0\u4e3a\u5341\u516d\u662f\u4e00\u4e2a\u975e\u5e38\u901a\u7528\u7684\u5e73\u8861\u503c\u3002<\/p>\n\n\n\n
    target_modules=[ \u2026 ]\uff1a\u6307\u5b9a\u8981\u5bf9\u6a21\u578b\u7684\u54ea\u4e9b\u7ebf\u6027\u5c42\uff08Linear Layers\uff09\u5e94\u7528 LoRA \u9002\u914d\u5668\u3002\u53ea\u5728\u8fd9\u4e9b\u5c42\u63d2\u5165\u53ef\u8bad\u7ec3\u7684 LoRA \u53c2\u6570\uff0c\u5176\u4ed6\u5c42\u4fdd\u6301\u51bb\u7ed3\u3002<\/p>\n\n\n\n
    lora_alpha=16\uff1aLoRA \u7684\u7f29\u653e\u7cfb\u6570\u7684\u7f29\u653e\u5206\u6bcd\uff0c\u7f29\u653e\u7cfb\u6570 = lora_alpha \/ r \u3002
    \u5982\u679c\u7f29\u653e\u7cfb\u6570 >1 \uff1a\u5fae\u8c03\u7684\u65b0\u77e5\u8bc6\u4f1a\u88ab\u653e\u5927\uff0c\u6a21\u578b\u53d8\u5316\u5267\u70c8\uff0c\u9002\u5e94\u65b0\u4efb\u52a1\u66f4\u5feb\uff0c\u4f46\u5bb9\u6613\u7834\u574f\u539f\u6709\u80fd\u529b\uff08\u707e\u96be\u6027\u9057\u5fd8\uff09\u3002
    \u5982\u679c\u7f29\u653e\u7cfb\u6570 =1 \uff1aLoRA \u7684\u5f71\u54cd\u6309\u539f\u59cb\u6bd4\u4f8b\u52a0\u5165\uff0c\u6700\u7a33\u5065\uff0c\u65e2\u5b66\u65b0\u4e1c\u897f\uff0c\u53c8\u4e0d\u600e\u4e48\u7834\u574f\u65e7\u77e5\u8bc6\u3002
    \u5982\u679c\u7f29\u653e\u7cfb\u6570 <1 \uff1a\u5fae\u8c03\u5f71\u54cd\u88ab\u524a\u5f31\uff0c\u8bad\u7ec3\u66f4\u5e73\u6ed1\u3001\u66f4\u4fdd\u5b88\uff0c\u9002\u5408\u6570\u636e\u91cf\u5f88\u5c0f\u6216\u9632\u6b62\u8fc7\u62df\u5408\u7684\u573a\u666f\u3002
    \u5f53\u524d\u7f29\u653e\u7cfb\u6570 = 16 \/ 16 = 1\uff0c\u6bd4\u8f83\u7a33\u59a5\u3002<\/p>\n\n\n\n
    lora_dropout=LORA_DROPOUT\uff1a\u5728\u8bad\u7ec3\u65f6\u968f\u673a\u4e22\u5f03\u4e00\u90e8\u5206\u795e\u7ecf\u5143\uff0c\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/p>\n\n\n\n
    bias=”none”\uff1a\u6307\u5b9a\u5982\u4f55\u5904\u7406\u6a21\u578b\u4e2d\u7684\u504f\u7f6e\u9879<\/strong>\uff08Bias\uff09\u53c2\u6570\uff0c”none”\uff08\u4e0d\u8bad\u7ec3\u4efb\u4f55\u504f\u7f6e\uff09\u3001”all”\uff08\u8bad\u7ec3\u6240\u6709\u504f\u7f6e\uff09\u3001”lora_only”\uff08\u53ea\u8bad\u7ec3 LoRA \u5c42\u5185\u65b0\u589e\u7684\u504f\u7f6e\uff09\uff0c\u5982\u679c\u8bbe\u7f6e\u4e3a\u8bad\u7ec3\u504f\u7f6e\u9879\uff0c\u9700\u8981\u6d88\u8017\u989d\u5916\u7684\u663e\u5b58\uff0c\u4f46\u5bf9\u6700\u7ec8\u4efb\u52a1\u7cbe\u5ea6\u51e0\u4e4e\u6ca1\u6709\u4f5c\u7528\u3002<\/p>\n\n\n\n
    use_gradient_checkpointing=”unsloth”\uff1a\u68af\u5ea6\u68c0\u67e5\u70b9\uff0c\u6709\u9009\u62e9\u5730\u4e22\u5f03\u5927\u90e8\u5206\u4e2d\u95f4\u6fc0\u6d3b\u503c\uff0c\u53ea\u4fdd\u7559\u5c11\u6570\u201c\u68c0\u67e5\u70b9\u201d\uff08Checkpoints\uff09\u3002\u5728\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u4ece\u6700\u8fd1\u7684\u68c0\u67e5\u70b9\u5f00\u59cb\uff0c\u91cd\u65b0\u8ba1\u7b97\u6240\u9700\u7684\u6fc0\u6d3b\u503c\u3002\u8fd9\u662f\u4e00\u79cd\u7ecf\u5178\u7684\u5185\u5b58\u4f18\u5316\u6280\u672f\uff0c\u80fd\u663e\u8457\u964d\u4f4e\u663e\u5b58\u5360\u7528\uff0c\u4f46\u4f1a\u56e0\u91cd\u65b0\u8ba1\u7b97\u800c\u589e\u52a0\u8bad\u7ec3\u65f6\u95f4\u3002\u8bbe\u7f6e\u4e3a”unsloth”\u65f6\uff0c\u80fd\u5728\u6781\u5c0f\u7684\u901f\u5ea6\u635f\u8017\u4e0b\uff08\u4ec5\u7ea6 +1.9% \u7684\u65f6\u95f4\u5f00\u9500\uff09\u6362\u6765\u4e86\u5de8\u5927\u7684\u663e\u5b58\u6536\u76ca\u3002<\/p>\n\n\n\n
    random_state=3407\uff1aLoRA \u6743\u91cd\u521d\u59cb\u5316\u7684\u968f\u673a\u79cd\u5b50\uff0c\u786e\u4fdd\u5b9e\u9a8c\u7ed3\u679c\u53ef\u590d\u73b0\u3002\u6bcf\u6b21\u8fd0\u884c\u4ee3\u7801\uff0cLoRA \u7684\u521d\u59cb\u72b6\u6001\u90fd\u4e00\u6837\u3002<\/p>\n\n\n\n
    use_rslora=False\uff1a\u662f\u5426\u4f7f\u7528 RS-LoRA\uff0c\u6807\u51c6 LoRA \u7f29\u653e\u662f alpha \/ r\uff0cRS-LoRA \u6539\u4e3a alpha \/ sqrt(r)\u3002\u5f53 r \u53d8\u5316\u65f6\uff0cRS-LoRA \u80fd\u4fdd\u6301\u66f4\u65b0\u91cf\u65b9\u5dee\u66f4\u7a33\u5b9a\uff0c\u4fbf\u4e8e\u8de8\u4e0d\u540c r \u503c\u8fdb\u884c\u8d85\u53c2\u8fc1\u79fb\u3002\u7531\u4e8e\u6211\u4eec\u7684\u9879\u76ee\u7f29\u653e\u4e3a1\uff0c\u56e0\u6b64\u5f00\u4e0d\u5f00\u542f\u6b64\u529f\u80fd\u5f71\u54cd\u4e0d\u5927\u3002<\/p>\n\n\n\n
    loftq_config=None\uff1a\u4e0d\u5f00\u542fLoFTQ\uff08Low-rank Factorization with Quantization\uff09\uff0c\u907f\u514d\u9879\u76ee\u521d\u59cb\u5316\u914d\u7f6e\u53d8\u5f97\u66f4\u52a0\u590d\u6742\u3002\u5982\u679c\u8981\u914d\u7f6eLoFTQ\uff0c\u5c31\u66f4\u52a0\u590d\u6742\u4e86\uff0c\u6211\u4eec\u5148\u4e0d\u7ba1\u8fd9\u91cc\u3002<\/p>\n\n\n\n
    \u8bb2\u89e3\u5b8c\u8fd9\u4e9b\u53c2\u6570\uff0c\u63a5\u4e0b\u6765\u6211\u4eec\u5f00\u59cb\u914d\u7f6eLoRA\uff1a<\/p>\n\n\n\n
        # ===================== STEP 2: LoRA =====================\n    log_section(\"STEP 2: \u914d\u7f6e LoRA\")\n\n    model = FastLanguageModel.get_peft_model(\n        model,\n        r=16,\n        target_modules=[\n            \"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n            \"gate_proj\", \"up_proj\", \"down_proj\",\n        ],\n        lora_alpha=16,\n        lora_dropout=LORA_DROPOUT,\n        bias=\"none\",\n        use_gradient_checkpointing=\"unsloth\",\n        random_state=3407,\n        use_rslora=False,\n        loftq_config=None,\n    )\n    log.info(\"\u2705 LoRA \u914d\u7f6e\u5b8c\u6210\")\n    model.print_trainable_parameters()<\/code><\/pre>\n\n\n\n
    7. \u51c6\u5907\u8bad\u7ec3\u96c6\u548c\u6837\u672c\u96c6<\/h3><\/div>\n\n\n\n
    \u63a5\u4e0b\u6765\u6211\u4eec\u628a\u8bad\u7ec3\u96c6\uff0c\u4e5f\u5c31\u662f\u6211\u4eec\u7684\u6837\u672c\u51c6\u5907\u597d\uff0c\u5148\u5199\u4e00\u4e2a\u65b9\u6cd5\u52a0\u8f7d\u8bad\u7ec3\u96c6\uff1a<\/p>\n\n\n\n
        # ===================== STEP 3: \u6570\u636e =====================\n    log_section(\"STEP 3: \u52a0\u8f7d\u6570\u636e\u96c6\")\n\n    def load_json_data(file_path):\n        with open(file_path, \"r\", encoding=\"utf-8\") as f:\n            return json.load(f)<\/code><\/pre>\n\n\n\n
    \u518d\u5199\u4e00\u4e2a\u65b9\u6cd5\uff0c\u628a\u8bad\u7ec3\u96c6\u4e2d\u7684\u6bcf\u4e00\u4e2a\u6837\u672c\u8f6c\u5316\u4e3a\u6a21\u578b\u8bad\u7ec3\u6240\u9700\u7684\u201c\u63d0\u793a\u8bcd\u6a21\u677f\u201d\uff1a<\/p>\n\n\n\n
        def convert_to_unsloth_format(example):\n        if all(k in example for k in [\"instruction\", \"input\", \"output\"]):\n            text = (\n                f\"### Instruction:\\n{example['instruction']}\\n\\n\"\n                f\"### Input:\\n{example['input']}\\n\\n\"\n                f\"### Response:\\n{example['output']}\"\n            )\n        elif all(k in example for k in [\"question\", \"answer\"]):\n            text = f\"### Question:\\n{example['question']}\\n\\n### Answer:\\n{example['answer']}\"\n        elif \"text\" in example:\n            text = example[\"text\"]\n        else:\n            text = str(example)\n        return {\"text\": text}<\/code><\/pre>\n\n\n\n
    \u63a5\u4e0b\u6765\u8c03\u7528\u8fd9\u4e24\u4e2a\u65b9\u6cd5\uff1a<\/p>\n\n\n\n
        raw_data = load_json_data(JSON_DATA_PATH)\n    log.info(f\"  \u539f\u59cb\u6761\u76ee\u6570: {len(raw_data)}\")\n    converted_data = [convert_to_unsloth_format(item) for item in raw_data]<\/code><\/pre>\n\n\n\n
    \u6211\u4eec\u518d\u7528\u4e00\u4e2a\u65b9\u6cd5\uff0c\u628aconverted_data\u8f6c\u5316\u4e3aHugging Face \u7684 Dataset \u5bf9\u8c61\u3002\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\u80fd\u5229\u7528 Hugging Face \u751f\u6001\u7684 map\u3001shuffle \u7b49\u9ad8\u6548\u6570\u636e\u5904\u7406\u65b9\u6cd5\u3002<\/p>\n\n\n\n
        dataset = hf_datasets.Dataset.from_list(converted_data)<\/code><\/pre>\n\n\n\n
    \u63a5\u4e0b\u6765\u6211\u4eec\u628a\u6837\u672c\u96c6\u62c6\u5206\u4e3a\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u96c6\u4e24\u4e2a\u90e8\u5206\uff1a<\/p>\n\n\n\n
        # train\/val \u62c6\u5206\n    split_dataset = dataset.train_test_split(test_size=VAL_SPLIT_RATIO, seed=3407)\n    train_dataset = split_dataset[\"train\"]\n    val_dataset = split_dataset[\"test\"]<\/code><\/pre>\n\n\n\n
    \u73b0\u5728\u8bad\u7ec3\u96c6\u548c\u6d4b\u8bd5\u9898\u5c31\u51c6\u5907\u5b8c\u6bd5\u4e86\uff0c\u6211\u4eec\u5199\u4e2a\u65e5\u5fd7\u8bb0\u5f55\u4e00\u4e0b\uff1a<\/p>\n\n\n\n
    log.info(f\"\u2705 \u6570\u636e\u96c6\u52a0\u8f7d\u5b8c\u6210\")
    log.info(f\"   \u8bad\u7ec3\u96c6: {len(train_dataset)} \u6761  |  \u9a8c\u8bc1\u96c6: {len(val_dataset)} \u6761\")
    sample = train_dataset[0][\"text\"][:200].replace(\"\\n\", \"\\\\n\")
    log.info(f\"   \u9996\u6761\u9884\u89c8: {sample}...\")<\/code><\/pre>\n\n\n\n
    \u8ba9\u6211\u4eec\u8fd0\u884c\u4e00\u4e0b\uff0c\u8bd5\u8bd5\u6548\u679c\uff0c\u4e3a\u4e86\u9875\u9762\u7684\u7b80\u6d01\uff0c\u7b14\u8005\u5c31\u53ea\u4fdd\u7559\u8fd9\u4e2a\u90e8\u5206\u7684\u65e5\u5fd7\u8f93\u51fa\u4e86\uff1a<\/p>\n\n\n\n
    16:44:45 | INFO    | ============================================================\n  STEP 3: \u52a0\u8f7d\u6570\u636e\u96c6\n============================================================\n16:44:45 | INFO    |   \u539f\u59cb\u6761\u76ee\u6570: 20036\n16:44:45 | INFO    | \u2705 \u6570\u636e\u96c6\u52a0\u8f7d\u5b8c\u6210\n16:44:45 | INFO    |    \u8bad\u7ec3\u96c6: 19034 \u6761  |  \u9a8c\u8bc1\u96c6: 1002 \u6761\n16:44:45 | INFO    |    \u9996\u6761\u9884\u89c8: ### Instruction:\\n\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5\uff0c\u8981\u6c42\u6587\u98ce\u4e00\u81f4\uff0c\u903b\u8f91\u8fde\u8d2f\u3002\\n\\n### Input:\\n\u7eb7\u7eb7\u5c0f\u5fc3\u8d77\u6765\u3002\\n\u3000\u3000\u4e0d\u8bf4\u8fd9\u5e7b\u9635\u662f\u5426\u96be\u7834\uff0c\u4e07\u4e00\u90a3\u4f4d\u9634\u7f57\u5b97\u7684\u5927\u957f\u8001\uff0c\u5728\u5e7b\u9635\u4e2d\u7a81\u7136\u6253\u5f00\u6740\u6212\uff0c\u4ed6\u4eec\u53c8\u6709\u8c01\u80fd\u591f\u5355\u72ec\u62b5\u6321\u3002\u81ea\u7136\u8fdf\u7591\u4e86\u8d77\u6765\u3002\\n\u3000\u3000\u770b\u5230\u8fd9\u79cd\u60c5\u5f62\uff0c\u82b1\u5929\u5947\u8138\u4e0a\u9732\u51fa\u4e00\u4e1d\u8ba5\u7b11\uff0c\u5f53\u5373\u4e0d\u518d\u7406\u4f1a\u5176\u4ed6\u4eba\uff0c\u4e00\u8f6c\u8eab\u51b2\u95e8\u4e0b\u5f1f\u5b50\u5429\u5490\u9053\uff1a\u201c\u4e0d\u7528\u4f60\u4eec\u5b88\u5728\u8fd9\u91cc\u4e86\u3002\u8fd9\u70b9...<\/code><\/pre>\n\n\n\n
    8. \u914d\u7f6e\u8bad\u7ec3<\/h3><\/div>\n\n\n\n
    \u63a5\u4e0b\u6765\u5c31\u662f\u91cd\u5934\u620f\u4e86\uff0c\u6211\u4eec\u9700\u8981\u914d\u7f6e\u8bad\u7ec3\u7684\u53c2\u6570\uff0c\u8fd9\u4e00\u6b65\u6bd4\u521a\u521a\u7684\u914d\u7f6e\u66f4\u52a0\u5355\u8c03\uff0c\u5982\u679c\u4e0d\u60f3\u770b\u7684\u8bdd\uff0c\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u6211\u914d\u7f6e\u597d\u7684\u914d\u7f6e\u9879\u3002<\/p>\n\n\n\n
    model=model\uff1a\u6ca1\u4ec0\u4e48\u597d\u8bf4\u7684\uff0c\u4f20\u5165\u57fa\u5ea7\u6a21\u578b\u3002<\/p>\n\n\n\n
    tokenizer=tokenizer\uff1a\u4f20\u5165\u5206\u8bcd\u5668\uff0c\u521a\u521a\u548c\u6a21\u578b\u4e00\u8d77\u62ff\u5230\u4e86\uff0c\u76f4\u63a5\u4f20\u5165\u3002<\/p>\n\n\n\n
    train_dataset=train_dataset\uff1a\u4f20\u5165\u8bad\u7ec3\u96c6\u3002<\/p>\n\n\n\n
    eval_dataset=val_dataset\uff1a\u4f20\u5165\u9a8c\u8bc1\u96c6\u3002\u8bad\u7ec3\u5668\u5728\u6bcf\u6b21\u8bc4\u4f30\u65f6\u8ba1\u7b97\u9a8c\u8bc1\u96c6\u635f\u5931\uff0c\u7528\u6765\u76d1\u63a7\u662f\u5426\u8fc7\u62df\u5408
    \u3002
    callbacks=[ClearCacheCache, EarlyStoppingCallback(\u2026)]\uff1aCallback \u56de\u8c03\u51fd\u6570\uff0c\u4f20\u5165ClearCacheCache\u65b9\u6cd5\uff0c\u6bcf\u6b21\u8bad\u7ec3\u6b65\u9aa4\u7ed3\u675f\u540e\u4e3b\u52a8\u6e05\u7406 GPU \u663e\u5b58\u788e\u7247\u3002EarlyStoppingCallback\u662f\u65e9\u505c\u7684\u65b9\u6cd5\uff0c\u6211\u4eec\u8bbe\u7f6e\u4e86EARLY_STOPPING_PATIENCE = 3\uff0c\u8fde\u7eed 3 \u6b21\u8bc4\u4f30\u90fd\u6ca1\u6709\u4e0b\u964d\uff0c\u5c31\u63d0\u524d\u7ec8\u6b62\u8bad\u7ec3\u3002\u53ef\u4ee5\u7701\u53bb\u65e0\u6548\u8bad\u7ec3\u65f6\u95f4\u3002<\/p>\n\n\n\n
    max_seq_length=MAX_SEQ_LENGTH\uff1a\u6a21\u578b\u80fd\u5904\u7406\u7684\u6700\u5927 Token \u957f\u5ea6\u3002\u6211\u4eec\u4e4b\u524d\u8bbe\u7f6e\u4e3a\u4e861280\u3002<\/p>\n\n\n\n
    per_device_train_batch_size=TRAIN_BATCH_SIZE\uff0c\u6bcf\u5f20 GPU \u4e0a\u7684\u6279\u5904\u7406\u5927\u5c0f\uff08Batch Size\uff09\u3002\u8fd9\u662f\u5b9e\u9645\u52a0\u8f7d\u5230\u5355\u5361\u663e\u5b58\u4e2d\u7684\u6837\u672c\u6570\u91cf\u3002\u7531\u4e8e\u6211\u4eec\u6709\u4e00\u5f20\u663e\u5361\uff0c\u56e0\u6b64\u4f20\u5165TRAIN_BATCH_SIZE\uff0c\u4e5f\u5c31\u662f\u4e4b\u524d\u8bbe\u7f6e\u597d\u76841\uff0c\u540c\u65f6\u5904\u7406\u4e00\u4e2a\u6837\u672c\u3002<\/p>\n\n\n\n
    gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS\uff1a\u68af\u5ea6\u7d2f\u79ef\u6b65\u6570\uff0c\u6211\u4eec\u4e4b\u524d\u8bbe\u7f6e\u4e3a\u4e864\uff0c\u6bcf4\u6b21\u66f4\u65b0\u4e00\u6b21\u6743\u91cd\u3002<\/p>\n\n\n\n
    warmup_steps=10\uff1a\u5b66\u4e60\u7387\u9884\u70ed\u6b65\u6570\u3002\u7531\u4e8e\u6211\u4eec\u662f\u505aLoRA\u5fae\u8c03\uff0c\u53ea\u9700\u8bbe\u7f6e\u8f83\u77ed\u7684\u9884\u70ed\u6b21\u6570\uff0c\u5982\u679c\u662f\u8fdb\u884c\u5168\u91cf\u5fae\u8c03\uff0c\u5219\u9700\u8981\u8c03\u9ad8\u8fd9\u4e2a\u503c\u3002<\/p>\n\n\n\n
    num_train_epochs=NUM_TRAIN_EPOCHS\uff1a\u8bad\u7ec3\u8f6e\u6570\uff0c\u6211\u4eec\u4e4b\u524d\u8bbe\u7f6e\u4e863\uff0c\u56e0\u6b64\u6837\u672c\u96c6\u6700\u591a\u53ef\u4ee5\u75283\u8f6e\u3002
    learning_rate=LEARNING_RATE\uff1a\u4f18\u5316\u5668\u7684\u5b66\u4e60\u7387\u3002LoRA \u5fae\u8c03\u901a\u5e38\u8bbe\u4e3a 1e-4 \u5230 5e-4 \u4e4b\u95f4\uff0c\u6211\u4eec\u4e4b\u524d\u8bbe\u7f6e\u4e3a\u4e862e-4\u3002<\/p>\n\n\n\n
    fp16=IS_V100\u548cbf16=not IS_V100\uff1a\u914d\u7f6e\u662f\u542f\u7528 BF16 \u8fd8\u662f FP16\u3002<\/p>\n\n\n\n
    logging_steps=10\uff1a\u6bcf\u9694 10 \u6b65\u5728\u63a7\u5236\u53f0\u6216\u65e5\u5fd7\u6587\u4ef6\u8f93\u51fa\u4e00\u6b21\u5f53\u524d\u7684 Loss \u548c\u5b66\u4e60\u7387\u3002<\/p>\n\n\n\n
    logging_dir=str(LOG_DIR)\uff1aTensorBoard \u6216\u65e5\u5fd7\u6587\u4ef6\u5b58\u653e\u7684\u76ee\u5f55\uff0c\u6211\u4eec\u8bbe\u7f6e\u8d77\u5c06\u65e5\u5fd7\u6587\u4ef6\u5b58\u653e\u5728\u672c\u5730\uff0c\u65b9\u4fbf\u65e5\u540e\u7684\u8c03\u8bd5\u3002<\/p>\n\n\n\n
    save_steps=SAVE_STEPS\uff1a\u6bcf\u9694\u591a\u5c11\u8bad\u7ec3\u6b65\u6570\u4fdd\u5b58\u4e00\u6b21\u68c0\u67e5\u70b9\u3002<\/p>\n\n\n\n
    save_total_limit=2\uff1a\u6700\u591a\u53ea\u4fdd\u7559\u6700\u8fd1\u7684 2 \u4e2a\u68c0\u67e5\u70b9\u3002\u65e7\u7684\u68c0\u67e5\u70b9\u4f1a\u88ab\u81ea\u52a8\u5220\u9664\uff0c\u8282\u7ea6\u786c\u76d8\u7a7a\u95f4\u3002<\/p>\n\n\n\n
    eval_steps=SAVE_STEPS\uff1a\u6bcf\u9694\u591a\u5c11\u6b65\u7528\u9a8c\u8bc1\u96c6\u505a\u4e00\u6b21\u9a8c\u8bc1\u3002\u8fd9\u91cc\u8bbe\u7f6e\u5f97\u548c save_steps \u5b8c\u5168\u4e00\u6837\uff0c\u610f\u5473\u7740\u6bcf\u6b21\u4fdd\u5b58\u65f6\u987a\u4fbf\u505a\u9a8c\u8bc1\u3002<\/p>\n\n\n\n
    save_strategy=”steps”\u548ceval_strategy=”steps”\uff1a\u4fdd\u5b58\u7b56\u7565\u548c\u9a8c\u8bc1\u7b56\u7565\u6309\u201c\u6b65\u6570\u201d\u6267\u884c\uff0c”no”\uff1a\u8bad\u7ec3\u671f\u95f4\u4e0d\u4fdd\u5b58\u548c\u9a8c\u8bc1\u4efb\u4f55\u4e2d\u95f4\u6a21\u578b\u3002\u53ea\u6709\u8bad\u7ec3\u7ed3\u675f\u624d\u6709\u6700\u7ec8\u6a21\u578b\u3002”epoch”\uff1a\u6bcf\u4e2a Epoch\uff08\u5b8c\u6574\u904d\u5386\u4e00\u6b21\u6570\u636e\u96c6\uff09\u7ed3\u675f\u540e\u4fdd\u5b58\u548c\u9a8c\u8bc1\u4e00\u6b21\u6a21\u578b\u3002”steps”\uff1a\u6bcf\u9694\u56fa\u5b9a\u7684 save_steps \u548c eval_strategy \u6b65\u6570\u4fdd\u5b58\u4e00\u6b21\u6a21\u578b\u3002<\/p>\n\n\n\n
    load_best_model_at_end=True\uff1a\u8bad\u7ec3\u7ed3\u675f\u540e\uff0c\u81ea\u52a8\u4ece\u6240\u6709\u68c0\u67e5\u70b9\u4e2d\u52a0\u8f7d\u9a8c\u8bc1\u96c6\u8868\u73b0\u6700\u597d\u7684\u90a3\u4e2a\u6a21\u578b\uff0c\u800c\u4e0d\u662f\u52a0\u8f7d\u6700\u540e\u4e00\u6b65\u7684\u6a21\u578b\u3002\u9632\u6b62\u4f60\u56e0\u4e3a\u540e\u9762\u8fc7\u62df\u5408\u800c\u9009\u5230\u5dee\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n
    metric_for_best_model=”eval_loss”\uff1a\u7528\u4ec0\u4e48\u6307\u6807\u5224\u5b9a\u201c\u6700\u597d\u201d\uff1f\u7528\u9a8c\u8bc1\u96c6\u635f\u5931\uff08eval_loss\uff09\u3002<\/p>\n\n\n\n
    greater_is_better=False\uff1a\u56e0\u4e3a\u635f\u5931\uff08Loss\uff09\u662f\u8d8a\u5c0f\u8d8a\u597d\uff0c\u6240\u4ee5\u586b False\u3002<\/p>\n\n\n\n
    output_dir=LORA_SAVE_DIR\uff1a\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u6743\u91cd\uff08LoRA \u9002\u914d\u5668\uff09\u548c\u4e2d\u95f4\u68c0\u67e5\u70b9\u4fdd\u5b58\u7684\u6587\u4ef6\u5939\u8def\u5f84\u3002
    optim=”adamw_8bit”\uff1a\u4f18\u5316\u5668\u7c7b\u578b\u3002adamw_8bit \u662f bitsandbytes \u5e93\u63d0\u4f9b\u7684 8-bit \u91cf\u5316\u7248 AdamW\uff0c\u80fd\u5c06\u4f18\u5316\u5668\u72b6\u6001\u5360\u7528\u7684\u663e\u5b58\u51cf\u5c11 75%\uff0c\u663e\u8457\u964d\u4f4e\u663e\u5b58\u7684\u4f7f\u7528\u3002<\/p>\n\n\n\n
    seed=3407\uff1a\u56fa\u5b9a\u6240\u6709\u968f\u673a\u64cd\u4f5c\u7684\u79cd\u5b50\uff0c\u786e\u4fdd\u6bcf\u6b21\u8fd0\u884c\u7ed3\u679c\u53ef\u590d\u73b0\u3002<\/p>\n\n\n\n
    dataset_num_proc=1\uff1a\u6570\u636e\u9884\u5904\u7406\u65f6\u4f7f\u7528\u7684 CPU \u8fdb\u7a0b\u6570\u3002\u5728\u5c06\u539f\u59cb\u6587\u672c\u8f6c\u6362\u6210\u6a21\u578b\u53ef\u7528\u7684 input_ids \u4e4b\u524d\uff0c\u542f\u52a8\u591a\u5c11\u4e2a\u72ec\u7acb\u7684CPU\u8fdb\u7a0b\u6765\u5e76\u884c\u5904\u7406\u6570\u636e\u96c6\u3002\u8bbe\u4e3a 1 \u662f\u4e3a\u4e86\u907f\u514d\u591a\u8fdb\u7a0b\u5e26\u6765\u7684\u6b7b\u9501\u6216\u5185\u5b58\u66b4\u589e\u95ee\u9898\u3002\u5efa\u8bae\u5f53\u6837\u672c\u91cf\u51e0\u5341\u4e07\u6761\u4ee5\u4e0a\uff0c\u518d\u8003\u8651\u628a\u8fd9\u4e2a\u503c\u8bbe\u7f6e\u4e3a2\u62164\uff0c\u5e76\u4e14\u8c28\u614e\u63d0\u9ad8\uff0c\u79ef\u6781\u8fdb\u884c\u6d4b\u8bd5\u3002<\/p>\n\n\n\n
    report_to=”none”\uff1a\u5173\u95ed\u5411\u5916\u90e8\u5b9e\u9a8c\u8ffd\u8e2a\u5e73\u53f0\u7684\u62a5\u544a\u3002\u8fd9\u91cc\u6211\u4eec\u4e0d\u5f04\u5f97\u592a\u590d\u6742\uff0c\u76f4\u63a5\u5173\u95ed\u3002<\/p>\n\n\n\n
    remove_unused_columns=True\uff1a\u81ea\u52a8\u5220\u9664\u6570\u636e\u96c6\u4e2d\u6a21\u578b\u4e0d\u9700\u8981\u7684\u5217\uff08\u5982\u539f\u59cb\u7684 text \u5b57\u6bb5\uff09\uff0c\u53ea\u4fdd\u7559 input_ids \u548c attention_mask \u4f20\u7ed9\u6a21\u578b\u3002\u5982\u679c\u4e0d\u5220\u9664\uff0c\u6a21\u578b\u4f1a\u56e0\u4e3a\u6536\u5230\u4e86\u591a\u4f59\u7684\u952e\u503c\u800c\u62a5\u9519\u3002\u5f53\u6211\u4eec\u628a\u5305\u542b {“text”: “### Instruction:\\n\u2026”} \u7684\u6570\u636e\u96c6\u4f20\u7ed9 SFTTrainer \u65f6\uff0c\u5b83\u5728\u6267\u884c trainer.train() \u4e4b\u524d\uff0c\u4f1a\u81ea\u52a8\u89e6\u53d1\u4e00\u4e2a\u5185\u90e8\u7684 _prepare_dataset() \u65b9\u6cd5\u3002\u8fd9\u4e2a\u65b9\u6cd5\u4f1a\u505a\u4e09\u4ef6\u4e8b\uff1a\u81ea\u52a8\u5206\u8bcd\uff08\u751f\u6210 input_ids\uff09\u3001\u81ea\u52a8\u751f\u6210\u63a9\u7801\uff08\u751f\u6210 attention_mask\uff09\u3001\u81ea\u52a8\u590d\u5236\u6807\u7b7e\uff08\u751f\u6210 labels\uff09\u3002\u5f53 SFTTrainer \u5728\u5185\u5b58\u4e2d\u751f\u6210\u4e86 input_ids, attention_mask, labels \u8fd9\u4e09\u4e2a\u65b0\u5217\u540e\uff0c\u6570\u636e\u96c6\u91cc\u540c\u65f6\u8fd8\u6709\u539f\u59cb\u7684 text \u5217\u3002\u6b64\u65f6\uff0cremove_unused_columns=True \u751f\u6548\uff0c\u5b83\u68c0\u67e5\u6a21\u578b\u7684 forward() \u7b7e\u540d\uff0c\u53d1\u73b0\u53ea\u9700\u8981 input_ids, attention_mask, labels\uff0c\u4e8e\u662f\u679c\u65ad\u5c06 text \u5217\u4ece\u6570\u636e\u6279\u6b21\u4e2d\u5254\u9664\u3002\u8fd9\u6837\uff0c\u6700\u7ec8\u8fdb\u5165 GPU \u7684\u6279\u6b21\u91cc\u5c31\u53ea\u6709\u5e72\u51c0\u7684\u6570\u503c\u5f20\u91cf\uff0c\u4e0d\u4f1a\u56e0\u4e3a\u591a\u4f59\u7684 text \u5b57\u7b26\u4e32\u800c\u62a5\u9519\u3002<\/p>\n\n\n\n
    dataloader_pin_memory=False\uff1a\u662f\u5426\u5c06\u6570\u636e\u9884\u52a0\u8f7d\u5230 CPU \u7684\u56fa\u5b9a\u5185\u5b58\uff08Page-locked Memory\uff09\u4ee5\u52a0\u901f\u4f20\u8f93\u3002\u7531\u4e8e\u6211\u4eec\u4e0d\u662f\u5728\u505a\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u6216\u8005\u8bad\u7ec3\u4e00\u4e2a\u89c4\u6a21\u5f88\u5c0f\u7684\u6a21\u578b\uff0c\u6570\u636e\u4f20\u8f93\u4e0d\u662f\u74f6\u9888\uff0c\u56e0\u6b64\u65e0\u9700\u8bbe\u7f6e\u4e3aTrue\uff0c\u5426\u5219\u4f1a\u989d\u5916\u5360\u7528\u5927\u91cf\u5185\u5b58\u3002<\/p>\n\n\n\n
    \u7ec8\u4e8e\u4ecb\u7ecd\u5b8c\u4e86\u8bad\u7ec3\u7684\u53c2\u6570\uff0c\u73b0\u5728\u6211\u4eec\u5c06\u8fd9\u4e9b\u53c2\u6570\u8fd0\u7528\u8d77\u6765\uff1a<\/p>\n\n\n\n
        # ===================== STEP 4: Trainer =====================\n    log_section(\"STEP 4: \u914d\u7f6e Trainer\")\n\n    trainer = SFTTrainer(\n        model=model,\n        tokenizer=tokenizer,\n        train_dataset=train_dataset,\n        eval_dataset=val_dataset,\n        callbacks=[ClearCacheCallback,\n                    EarlyStoppingCallback(early_stopping_patience=EARLY_STOPPING_PATIENCE)],\n        args=SFTConfig(\n            max_seq_length=MAX_SEQ_LENGTH,\n            per_device_train_batch_size=TRAIN_BATCH_SIZE,\n            gradient_accumulation_steps=GRADIENT_ACCUMULATION_STEPS,\n            warmup_steps=10,\n            num_train_epochs=NUM_TRAIN_EPOCHS,\n            learning_rate=LEARNING_RATE,\n            fp16=IS_V100,          # V100 \u7528 fp16\uff0c\u5176\u4ed6\u5173\n            bf16=not IS_V100,      # \u975e V100 \u7528 bf16\n            logging_steps=10,\n            logging_dir=str(LOG_DIR),\n            save_steps=SAVE_STEPS,\n            save_total_limit=2,\n            eval_steps=SAVE_STEPS,     # \u4e0e save \u540c\u9891\u505a\u9a8c\u8bc1\n            save_strategy=\"steps\",\n            eval_strategy=\"steps\",\n            load_best_model_at_end=True,\n            metric_for_best_model=\"eval_loss\",\n            greater_is_better=False,\n            output_dir=LORA_SAVE_DIR,\n            optim=\"adamw_8bit\",\n            seed=3407,\n            dataset_num_proc=1,\n            report_to=\"none\",\n            remove_unused_columns=True,\n            dataloader_pin_memory=False,\n        ),\n    )\n    log.info(\"\u2705 Trainer \u914d\u7f6e\u5b8c\u6210\")\n    log.info(f\"   \u7cbe\u5ea6: {'fp16' if IS_V100 else 'bf16'}  |  eval \u6bcf {SAVE_STEPS} steps\")<\/code><\/pre>\n\n\n\n
    9. \u5f00\u59cb\u8bad\u7ec3<\/h3><\/div>\n\n\n\n
    \u51c6\u5907\u5de5\u4f5c\u90fd\u505a\u5b8c\u4e86\uff0c\u73b0\u5728\u6211\u4eec\u5199\u5f00\u59cb\u8fd0\u884c\u8bad\u7ec3\u7684\u4ee3\u7801\uff0c\u5148\u5199\u4e00\u4e9b\u65e5\u5fd7\uff1a<\/p>\n\n\n\n
        # ===================== STEP 5: \u8bad\u7ec3 =====================\n    log_section(\"STEP 5: \u5f00\u59cb\u8bad\u7ec3\")\n    eff_batch = TRAIN_BATCH_SIZE * GRADIENT_ACCUMULATION_STEPS\n    log.info(f\"  GPU={gpu_name}  VRAM={vram_gb:.0f}GB  dtype={'fp16' if IS_V100 else 'bf16'}\")\n    log.info(f\"  epochs={NUM_TRAIN_EPOCHS}  batch_size={TRAIN_BATCH_SIZE}  grad_accum={GRADIENT_ACCUMULATION_STEPS}  effective_batch={eff_batch}\")\n    log.info(f\"  max_seq_length={MAX_SEQ_LENGTH}  lr={LEARNING_RATE}\")<\/code><\/pre>\n\n\n\n
    \u73b0\u5728\u6211\u4eec\u6765\u5199\u4e00\u4e9b\u4ee3\u7801\uff0c\u7528\u6765\u5b9e\u73b0\u65ad\u70b9\u540e\u542f\u52a8\uff0c\u7ee7\u7eed\u8bad\u7ec3\u3002\u56e0\u4e3a\u6211\u4eec\u521a\u521a\u914d\u7f6e\u4e86\u68c0\u67e5\u70b9\uff0c\u56e0\u6b64\u6211\u4eec\u53ea\u9700\u8981\u53bb\u627e\u6709\u6ca1\u6709\u68c0\u67e5\u70b9\u6587\u4ef6\uff0c\u5c31\u53ef\u4ee5\u77e5\u9053\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u8fd8\u662f\u4ece\u68c0\u67e5\u70b9\u7ee7\u7eed\u8bad\u7ec3\u4e86\uff1a<\/p>\n\n\n\n
        resume = False\n    if Path(LORA_SAVE_DIR).exists():\n        checkpoints = sorted(Path(LORA_SAVE_DIR).glob(\"checkpoint-*\"))\n        if checkpoints:\n            resume = True\n            log.info(f\"\ud83d\udcc2 \u4ece checkpoint \u6062\u590d: {checkpoints[-1].name}\")\n    if not resume:\n        log.info(\"\ud83c\udd95 \u9996\u6b21\u8bad\u7ec3\uff0c\u4ece\u5934\u5f00\u59cb\")<\/code><\/pre>\n\n\n\n
    \u6700\u540e\u518d\u5199\u4e00\u4e2atrainer.train()\u65b9\u6cd5\uff0c\u5f00\u542f\u8bad\u7ec3\uff1a<\/p>\n\n\n\n
        t_train = time.time()\n    trainer.train(resume_from_checkpoint=resume)\n    log.info(f\"\u2705 \u8bad\u7ec3\u5b8c\u6210\uff01\u603b\u8017\u65f6 {(time.time() - t_train) \/ 60:.1f} \u5206\u949f\")<\/code><\/pre>\n\n\n\n
    \u8ba9\u6211\u4eec\u8dd1\u8d77\u6765\u770b\u770b\u6548\u679c\uff1a<\/p>\n\n\n\n
    02:15:28 | ERROR   | \ud83d\udca5 \u811a\u672c\u5f02\u5e38\u9000\u51fa\uff0c\u8be6\u89c1\u65e5\u5fd7: D:\\Data\\Python\\qwopus3.5-9b-fanren-lora\\logs\\train_20260726_021306.log<\/code><\/pre>\n\n\n\n
    \u5947\u602a\uff0c\u5c45\u7136\u62a5\u9519\u4e86\uff0c\u8fd9\u662f\u600e\u4e48\u56de\u4e8b\uff1f\u5728\u7b14\u8005\u4e4b\u524d\u7684\u6d4b\u8bd5\u4e2d\uff0c\u8fd9\u4e2a\u4ee3\u7801\u662f\u80fd\u6b63\u5e38\u8fd0\u884c\uff0c\u5e76\u4e14\u5df2\u7ecf\u628a\u6a21\u578b\u8dd1\u51fa\u6765\u4e86\u7684\uff0c\u4e3a\u4ec0\u4e48\u8fd9\u4e2a\u65f6\u5019\u7a81\u7136\u62a5\u9519\u4e86\u5462\uff1f<\/p>\n\n\n\n
    10. \u5947\u602a\u7684\u62a5\u9519<\/h3><\/div>\n\n\n\n
    \u62a5\u9519\u6574\u4f53\u6bd4\u8f83\u957f\u3002\u8fd9\u91cc\u7b14\u8005\u6c42\u52a9\u4e86Deepseek\u5206\u6790\u8fd9\u4e2a\u95ee\u9898\uff0c\u5927\u6982\u539f\u56e0\u662fUnsloth \u4eceHuggingFace \u62ff\u6a21\u578b\u7684\u65f6\u5019\uff0c\u65b9\u6cd5\u7684\u5bf9\u63a5\u51fa\u4e86\u4e00\u4e2a\u95ee\u9898\uff0c\u62ff\u4e0d\u5230*args\u53d8\u91cf\uff0c\u7531\u4e8e\u8fd9\u91cc\u4e0d\u662f\u672c\u6587\u7684\u91cd\u70b9\uff0c\u56e0\u6b64\u5c31\u4e0d\u5728\u8fd9\u91cc\u82b1\u66f4\u591a\u7684\u65f6\u95f4\u9610\u8ff0\u4e86\uff0c\u7b14\u8005\u8ba9Deepseek\u5199\u4e86\u4e00\u6bb5\u4ee3\u7801\u4f5c\u4e3a\u8865\u4e01\uff0c\u653e\u5728\u5bfc\u5165\u5305\u548c\u5168\u5c40\u914d\u7f6e\u4e2d\u95f4\uff1a<\/p>\n\n\n\n
    # \u2500\u2500 Monkey-patch: \u4fee\u590d Unsloth \u7f16\u8bd1\u5668\u4e0e force_accelerate_hooks \u7684\u517c\u5bb9\u6027 \u2500\u2500\n# force_accelerate_hooks \u628a forward \u66ff\u6362\u4e3a wrapped(*args, **kwargs)\uff0c\u4f46\u6ca1\u6709\u8bbe\n# __wrapped__ \u5c5e\u6027\uff0c\u5bfc\u81f4 inspect.signature \u4e22\u5931\u539f\u59cb\u663e\u5f0f\u53c2\u6570\u540d\uff08\u53d8\u6210 *args\uff09\u3002\n# Unsloth \u7f16\u8bd1\u5668\u7528\u8fd9\u4e2a\u9519\u8bef\u7684\u7b7e\u540d\u751f\u6210\u7f16\u8bd1\u7f13\u5b58\uff0c\u4ea7\u51fa\u7684 forward \u65b9\u6cd5\u7b7e\u540d\u5934\u4f7f\u7528\n# \u663e\u5f0f\u53c2\u6570\u540d\u800c\u8c03\u7528\u4f53\u5199 *args\uff0c\u5f15\u53d1 NameError\u3002\n# \u672c\u8865\u4e01\u4ec5\u6dfb\u52a0 __wrapped__ \u5f15\u7528\uff0c\u4e0d\u6539\u53d8 wrapper \u7684\u4efb\u4f55\u8fd0\u884c\u65f6\u884c\u4e3a\u3002\nimport transformers.integrations.accelerate as _hf_accelerate\n_original_force_accelerate_hooks = _hf_accelerate.force_accelerate_hooks\ndef _patched_force_accelerate_hooks(child_module_name):\n    _original_decorator = _original_force_accelerate_hooks(child_module_name)\n    def _decorator(forward_func):\n        wrapped_fn = _original_decorator(forward_func)\n        wrapped_fn.__wrapped__ = forward_func  # \u8ba9 inspect.signature \u7a7f\u900f\u5230\u539f\u59cb\u7b7e\u540d\n        return wrapped_fn\n    return _decorator\n_hf_accelerate.force_accelerate_hooks = _patched_force_accelerate_hooks\n# \u2500\u2500 \u8865\u4e01\u7ed3\u675f \u2500\u2500<\/code><\/pre>\n\n\n\n
    11. \u5f00\u59cb\u8bad\u7ec3<\/h3><\/div>\n\n\n\n
    \u6211\u4eec\u518d\u4e00\u6b21\u542f\u52a8\uff0c\u5148\u8981\u6c42powershell\u8fdb\u5165UTF-8\u7f16\u7801\uff0c\u518d\u5728\u8fd0\u884c\u811a\u672c\uff0c\u53ef\u4ee5\u51cf\u5c11\u62a5\u9519\uff1a<\/p>\n\n\n\n
    (lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>chcp 65001\nActive code page: 65001\n\n(lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>python .\/train.py<\/code><\/pre>\n\n\n\n
    \u5927\u6982\u8fd0\u884c\u4e86\u4e24\u4e09\u5206\u949f\uff0c\u80fd\u770b\u5230\u5df2\u7ecf\u5728\u8f93\u51fa\u8fdb\u5ea6\u4e86\uff1a<\/p>\n\n\n\n
    (lora) D:\\Data\\Python\\qwopus3.5-9b-fanren-lora>Unsloth: Double buffering enabled (parallel H2D + compute) for backward pass.\n  0%|                                                                             | 6\/14277 [00:58<19:26:32,  4.90s\/it]<\/code><\/pre>\n\n\n\n
    \u521a\u5f00\u59cb\u8bad\u7ec3\u901f\u5ea6\u4f1a\u7a0d\u5fae\u8f83\u6162\uff0c\u4e4b\u540e\u4f1a\u8fbe\u5230\u6bd4\u8f83\u5e73\u7a33\u7684\u901f\u5ea6\u3002\u7531\u4e8e\u6700\u7ec8\u5927\u6982\u7387\u4f1a\u89e6\u53d1\u65e9\u505c\uff0c\u56e0\u6b64\u5b9e\u9645\u8bad\u7ec3\u65f6\u95f4\u4f1a\u6bd4\u9884\u8ba1\u8bad\u7ec3\u65f6\u95f4\u7565\u77ed\u4e00\u4e9b\u3002\u4f46\u5148\u522b\u6025\u7740\u8bad\u7ec3\uff0c\u6211\u4eec\u5148\u628a\u4fdd\u5b58\u6a21\u578b\u7684\u4ee3\u7801\u5199\u597d\u3002<\/p>\n\n\n\n
    12. \u4fdd\u5b58\u6a21\u578b<\/h3><\/div>\n\n\n\n
    \u7531\u4e8e\u6211\u4eec\u53ea\u662f\u5bf9\u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u8c03\uff0c\u56e0\u6b64\u53ea\u9700\u8981\u4fdd\u5b58\u5fae\u8c03\u4ea7\u751f\u7684LoRA \u8f7b\u91cf\u5316\u6743\u91cd\u548c\u5206\u8bcd\u5668\u3002\u81f3\u4e8e\u4e3a\u4ec0\u4e48\u9700\u8981\u4fdd\u5b58\u5206\u8bcd\u5668\uff0c\u53ea\u662f\u56e0\u4e3aHugging Face\u63a8\u8350\u6211\u4eec\u8fd9\u4e48\u505a\uff0c\u90a3\u6211\u4eec\u5c31\u987a\u624b\u4fdd\u5b58\u4e86\uff0c\u5e76\u4e14\u5728\u4e0a\u4f20\u6211\u4eec\u7684\u5fae\u8c03\u6a21\u578b\u7684\u65f6\u5019\u4e5f\u987a\u624b\u628a\u5b83\u4e0a\u4f20\u4e0a\u53bb\uff0c\u5b9e\u9645\u4e0a\u4f5c\u7528\u4e0d\u5927\uff0c\u6211\u4eec\u5728\u63a5\u4e0b\u6765\u7684\u8f93\u51fa\u811a\u672c\u4e2d\u4e5f\u4e0d\u4f1a\u4f7f\u7528\u5230\u3002<\/p>\n\n\n\n
    \u56e0\u6b64\uff0c\u6211\u4eec\u6765\u5b8c\u6210\u4fdd\u5b58\u6a21\u578b\u90e8\u5206\u7684\u4ee3\u7801\uff1a<\/p>\n\n\n\n
        # ===================== STEP 6: \u4fdd\u5b58 =====================\n    log_section(\"STEP 6: \u4fdd\u5b58\u6a21\u578b\")\n    # \u4fdd\u5b58LoRA \u8f7b\u91cf\u5316\u6743\u91cd\n    model.save_pretrained(LORA_SAVE_DIR)\n    # \u4fdd\u5b58\u5206\u8bcd\u5668\n    tokenizer.save_pretrained(LORA_SAVE_DIR)\n    log.info(f\"\ud83c\udf89 LoRA \u9002\u914d\u5668\u5df2\u4fdd\u5b58\u81f3: {LORA_SAVE_DIR}\")\n    log.info(f\"   \u5b8c\u6574\u65e5\u5fd7: {LOG_FILE.resolve()}\")<\/code><\/pre>\n\n\n\n
    13. \u8fd0\u884c\u5165\u53e3<\/h3><\/div>\n\n\n\n
    \u6700\u540e\uff0c\u6211\u4eec\u6765\u5199\u4e00\u4e2a\u8fd0\u884c\u5165\u53e3\uff0c\u5e76\u4e14\u8f93\u51fa\u4e00\u4e9b\u65e5\u5fd7\uff0c\u65b9\u4fbf\u6211\u4eec\u7684\u8c03\u8bd5\uff1a<\/p>\n\n\n\n
    # ===================== \u5165\u53e3 =====================\nif __name__ == \"__main__\":\n    log_section(\"\u73af\u5883\u8bca\u65ad\")\n    try:\n        main()\n    except Exception:\n        log.error(\"\\n\" + traceback.format_exc())\n        log.error(f\"\ud83d\udca5 \u811a\u672c\u5f02\u5e38\u9000\u51fa\uff0c\u8be6\u89c1\u65e5\u5fd7: {LOG_FILE.resolve()}\")\n        sys.exit(1)<\/code><\/pre>\n\n\n\n
    14. \u5f00\u59cb\u8bad\u7ec3<\/h3><\/div>\n\n\n\n
    \u63a5\u4e0b\u6765\u5c31\u5230\u4e86\u6700\u8017\u8d39\u65f6\u95f4\u7684\u6b65\u9aa4\u3002\u8fd0\u884c\u6211\u4eec\u7684train.py\uff0c\u5f00\u59cb\u6211\u4eec\u7684\u8bad\u7ec3\uff0c\u8fd9\u4e2a\u6b65\u9aa4\u53ef\u80fd\u6d88\u80171~2\u5929\u3002\u5f53\u7136\u8bfb\u8005\u4eec\u4e5f\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u7b14\u8005\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684\u6a21\u578b<\/a>\uff0c\u8282\u7ea6\u65f6\u95f4\u548c\u7535\u8d39~<\/p>\n\n\n\n
    \u7b2c\u4e09\u6b65\u2014\u2014\u7f16\u5199\u8fd0\u884c\u6a21\u578b\u7684\u811a\u672c<\/h2><\/div>\n\n\n\n
    \u957f\u8fbe\u4e24\u5929\u7684\u65f6\u95f4\uff0c\u6211\u4eec\u7684\u7535\u8111\u6d88\u8017\u4e865~6\u5757\u7684\u7535\u8d39\u4e4b\u540e\uff0c\u7ec8\u4e8e\u5c06\u5fae\u8c03\u8dd1\u51fa\u6765\u4e86\u3002\u7531\u4e8e\u6211\u4eec\u8bbe\u7f6e\u4e86LORA_SAVE_DIR    = \".\/lora\/qwen3.5-9b-text-lora-checkpoint\"<\/code>\uff0c\u56e0\u6b64\u5f97\u53bb.\/lora\/qwen3.5-9b-text-lora-checkpoint<\/code><\/code>\u6587\u4ef6\u5939\u4e0b\u627e\u5230\u6211\u4eec\u7684\u5fae\u8c03\u6587\u4ef6\uff0c\u90a3\u4e48\u5982\u4f55\u4f7f\u7528\u6211\u4eec\u7684\u5fae\u8c03\u6a21\u578b\u8fdb\u884c\u5199\u4f5c\u5462\uff1f\u6211\u4eec\u9700\u8981\u5b9e\u73b0\u4e00\u4e2a\u7b80\u5355\u7684\u8f93\u51fa\u811a\u672c\uff0c\u4e5f\u5c31\u662f\u9879\u76ee\u4e2d\u7684inference.py\u3002\u9879\u76ee\u4e2d\u7684inference.py\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5185\u5d4c\u4e86\u4e00\u4e9b\u8c03\u8bd5\u7528\u7684\u9884\u8bbe\u6587\u672c\u3002\u6211\u4eec\u73b0\u5728\u4e0d\u9700\u8981\u5b9e\u73b0\u8fd9\u4e9b\u3002<\/p>\n\n\n\n
    1. \u5bfc\u5165\u5305<\/h3><\/div>\n\n\n\n
    \u6309\u7167\u60ef\u4f8b\uff0c\u5148\u5bfc\u4e2a\u5305\uff0c\u5e76\u4e14\u8981\u6c42\u7f16\u7801\u4e3aUTF-8:<\/p>\n\n\n\n
    import sys\nimport os\n\n# Windows UTF-8 \u6a21\u5f0f\u5f3a\u5236\uff08\u540c train.py\uff09\nif sys.flags.utf8_mode == 0:\n    os.environ[\"PYTHONUTF8\"] = \"1\"\n    os.execv(sys.executable, [sys.executable] + sys.argv)\n\nimport argparse\nimport torch\nimport time\nfrom pathlib import Path\nfrom unsloth import FastLanguageModel\nfrom transformers import TextStreamer<\/code><\/pre>\n\n\n\n
    2. \u6253\u4e2a\u8865\u4e01<\/h3><\/div>\n\n\n\n
    \u4e4b\u540e\u548ctrain.py\u4e00\u6837\uff0c\u4e3aUnsloth \u4eceHuggingFace \u6253\u4e00\u4e2a\u8865\u4e01\uff0c\u8fd9\u4e2a\u5730\u65b9\u6211\u4eec\u6682\u65f6\u4e0d\u53bb\u7406\u89e3\uff1a<\/p>\n\n\n\n
    # \u2500\u2500 Monkey-patch: \u4fee\u590d Unsloth \u7f16\u8bd1\u5668\u4e0e force_accelerate_hooks \u7684\u517c\u5bb9\u6027 \u2500\u2500\n# \u8be6\u89c1 train.py \u4e2d\u540c\u540d\u8865\u4e01\u7684\u6ce8\u91ca\u3002\nimport transformers.integrations.accelerate as _hf_accelerate\n_original_force_accelerate_hooks = _hf_accelerate.force_accelerate_hooks\ndef _patched_force_accelerate_hooks(child_module_name):\n    _original_decorator = _original_force_accelerate_hooks(child_module_name)\n    def _decorator(forward_func):\n        wrapped_fn = _original_decorator(forward_func)\n        wrapped_fn.__wrapped__ = forward_func\n        return wrapped_fn\n    return _decorator\n_hf_accelerate.force_accelerate_hooks = _patched_force_accelerate_hooks\n# \u2500\u2500 \u8865\u4e01\u7ed3\u675f \u2500\u2500<\/code><\/pre>\n\n\n\n
    3. \u8bbe\u7f6e\u968f\u673a\u79cd\u5b50<\/h3><\/div>\n\n\n\n
    \u5bf9\u4e8e\u8f93\u51fa\uff0c\u6211\u4eec\u80af\u5b9a\u5e0c\u671b\u6a21\u578b\u7684\u6bcf\u6b21\u8f93\u51fa\u90fd\u662f\u4e0d\u4e00\u6837\u7684\uff0c\u56e0\u6b64\u6211\u4eec\u8bbe\u7f6e\u5176\u79cd\u5b50\u503c\u662f\u968f\u673a\u7684\uff1a<\/p>\n\n\n\n
    # \u968f\u673a\u79cd\u5b50\uff1a\u6bcf\u6b21\u8fd0\u884c\u751f\u6210\u4e0d\u540c\u7ed3\u679c\n# \u8bbe\u7f6ePyTorch\u5728CPU\u4e0a\u7684\u968f\u673a\u64cd\u4f5c\u7684\u79cd\u5b50\u503c\ntorch.manual_seed(int(time.time() * 1000) % (2**31))\n# \u8bbe\u7f6ePyTorch\u5728GPU\u4e0a\u7684\u968f\u673a\u64cd\u4f5c\u7684\u79cd\u5b50\u503c\nif torch.cuda.is_available():\n    torch.cuda.manual_seed_all(int(time.time() * 1000) % (2**31))<\/code><\/pre>\n\n\n\n
    4. \u8bbe\u7f6e\u57fa\u5ea7\u6a21\u578b<\/h3><\/div>\n\n\n\n
    \u548ctrain.py\u4e00\u6837\uff0c\u6211\u4eec\u9700\u8981\u8bbe\u7f6e\u57fa\u5ea7\u6a21\u578b\u7684\u4f4d\u7f6e\uff0c\u5982\u679c\u6ca1\u6709\u7684\u8bdd\u81ea\u52a8\u4eceHugging Face\u4e0a\u4e0b\u8f7d\uff0c\u5982\u679c\u6709\u7684\u8bdd\u5219\u76f4\u63a5\u4f7f\u7528\u672c\u5730\u6a21\u578b\uff1a<\/p>\n\n\n\n
    # \u4f18\u5148\u4f7f\u7528\u672c\u5730\u6a21\u578b\uff0c\u4e0d\u5b58\u5728\u5219\u4ece HuggingFace \u81ea\u52a8\u4e0b\u8f7d\nLOCAL_MODEL_DIR = str(Path(__file__).resolve().parent \/ \"model\/Qwen3.5-9B-Text-Only-abliterated\")\nHF_MODEL_ID = \"zjml\/Qwen3.5-9B-Text-Only-abliterated\"\nMODEL_ID = LOCAL_MODEL_DIR if Path(LOCAL_MODEL_DIR).exists() else HF_MODEL_ID\n\ndef load_model(lora_path: str):\n    \"\"\"\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b + LoRA \u9002\u914d\u5668\"\"\"\n    print(f\"\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b: {MODEL_ID}\")\n    model, tokenizer = FastLanguageModel.from_pretrained(\n        model_name=MODEL_ID,\n        max_seq_length=1280,\n        dtype=torch.bfloat16,\n        load_in_4bit=True,\n    )\n    print(f\"\u52a0\u8f7d LoRA \u9002\u914d\u5668: {lora_path}\")\n    model.load_adapter(lora_path)\n    model.to(\"cuda\")  # \u786e\u4fdd LoRA \u6743\u91cd\u79fb\u5230 GPU\n    FastLanguageModel.for_inference(model)\n    print(\"\u2705 \u6a21\u578b\u5c31\u7eea\")\n    return model, tokenizer<\/code><\/pre>\n\n\n\n
    5. \u64b0\u5199\u63d0\u793a\u8bcd\u6a21\u677f<\/h3><\/div>\n\n\n\n
    \u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u6765\u5199\u4e00\u6bb5\u63d0\u793a\u8bcd\u6a21\u677f\u3002\u8fd8\u662f\u8001\u751f\u5e38\u8c08\u7684\u95ee\u9898\uff0c\u6211\u4eec\u5728\u8bad\u7ec3\u7684\u65f6\u5019\u66f4\u63a8\u8350\u5176\u5b9e\u662f\u4f7f\u7528\u683c\u5f0f 2\uff1aquestion-answer\uff0c\u8fd9\u6837\u5c31\u4e0d\u9700\u8981\u6709Instruction\u90e8\u5206\u5360\u7528\u4e0a\u4e0b\u6587\u4e86\u3002\u4e0d\u8fc7\u65e2\u7136\u6a21\u578b\u5df2\u7ecf\u8bad\u7ec3\u51fa\u6765\u4e86\uff0c\u6211\u4eec\u5c31\u7528\u683c\u5f0f1\u6765\u5199\u4e00\u6bb5\u63d0\u793a\u8bcd\u6a21\u677f\uff0c\u65b9\u4fbf\u4e4b\u540e\u7684\u4ee3\u7801\u4e2d\u76f4\u63a5\u66ff\u6362input\u90e8\u5206\u6765\u5b9e\u73b0\u4e0a\u4e0b\u6587\u7684\u7eed\u5199\uff1a<\/p>\n\n\n\n
    PROMPT_TEMPLATE = \"\"\"### Instruction:\n\u4f60\u662f\u4ed9\u4fa0\u5c0f\u8bf4\u300a\u51e1\u4eba\u4fee\u4ed9\u4f20\u300b\u98ce\u683c\u7684\u4f5c\u5bb6\uff0c\u8bf7\u6839\u636e\u4ee5\u4e0b\u524d\u6587\u5185\u5bb9\uff0c\u7eed\u5199\u63a5\u4e0b\u6765\u7684\u5267\u60c5\uff0c\u8981\u6c42\u6587\u98ce\u4e00\u81f4\uff0c\u903b\u8f91\u8fde\u8d2f\u3002\n\n### Input:\n{input_text}\n\n### Response:\n\"\"\"<\/code><\/pre>\n\n\n\n
    6. \u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b\u548c\u5206\u8bcd\u5668<\/h3><\/div>\n\n\n\n
    \u73b0\u5728\uff0c\u6211\u4eec\u5199\u4e00\u4e2a\u65b9\u6cd5\u6765\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b\u548c\u5206\u8bcd\u5668\uff0c\u5148\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b\uff0c\u518d\u52a0\u8f7dLoRA\u9002\u914d\u5668\uff0c\u4e4b\u540e\u786e\u4fdd\u6574\u4e2a\u7ec4\u5408\u6a21\u578b\uff08\u57fa\u5ea7 + LoRA\uff09\u5b8c\u6574\u5730\u642c\u8fd0\u5230 GPU \u663e\u5b58\uff0c\u6700\u540e\u5f00\u542f\u63a8\u7406\u4f18\u5316\u6a21\u5f0f\uff1a<\/p>\n\n\n\n
    def load_model(lora_path: str):\n    \"\"\"\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b + LoRA \u9002\u914d\u5668\"\"\"\n    print(f\"\u52a0\u8f7d\u57fa\u5ea7\u6a21\u578b: {MODEL_ID}\")\n    model, tokenizer = FastLanguageModel.from_pretrained(\n        model_name=MODEL_ID,\n        max_seq_length=1280,\n        dtype=torch.bfloat16,\n        load_in_4bit=True,\n    )\n    print(f\"\u52a0\u8f7d LoRA \u9002\u914d\u5668: {lora_path}\")\n    model.load_adapter(lora_path)\n    model.to(\"cuda\")  # \u786e\u4fdd LoRA \u6743\u91cd\u79fb\u5230 GPU\n    # \u5f00\u542f\u63a8\u7406\u4f18\u5316\u6a21\u5f0f\n    FastLanguageModel.for_inference(model)\n    print(\"\u2705 \u6a21\u578b\u5c31\u7eea\")\n    return model, tokenizer<\/code><\/pre>\n\n\n\n
    7. \u8f93\u51fa\u65b9\u6cd5<\/h3><\/div>\n\n\n\n
    \u73b0\u5728\u6211\u4eec\u5199\u4e00\u4e2a\u65b9\u6cd5\u6765\u5b9e\u73b0\u6d41\u5f0f\u8f93\u51fa\u3002\u5bf9\u4e8e\u8bed\u8a00\u6a21\u578b\uff0c\u6211\u4eec\u9996\u5148\u5148\u8981\u5c06\u8f93\u5165\u7684\u6587\u672c\u901a\u8fc7\u5206\u8bcd\u5668\u8f6c\u5316\u4e3a\u6a21\u578b\u80fd\u591f\u8bfb\u61c2\u7684\u8bed\u8a00\uff1a<\/p>\n\n\n\n
    def generate(model, tokenizer, prompt: str, max_tokens: int, temperature: float, top_p: float):\n    \"\"\"\u6d41\u5f0f\u751f\u6210\u6587\u672c\"\"\"\n    inputs = tokenizer(prompt, return_tensors=\"pt\").to(\"cuda\")<\/code><\/pre>\n\n\n\n
    \u7136\u540e\u6253\u5f00\u6d41\u5f0f\u8f93\u51fa\u5f00\u5173\uff1a<\/p>\n\n\n\n
        streamer = TextStreamer(tokenizer, skip_prompt=True)<\/code><\/pre>\n\n\n\n
    \u63a5\u4e0b\u6765\u5c31\u662f\u8c03\u7528model.generate()\u65b9\u6cd5\uff0c\u5e76\u4e14\u628a\u8bf8\u5982\u6e29\u5ea6\u7b49\u6570\u503c\u4f20\u5165\u8fdb\u53bb\uff0c\u7531\u4e8e\u8fd9\u4e2a\u65b9\u6cd5\u5728\u6267\u884c\u4e2d\u4f1a\u81ea\u5df1\u6253\u5370\u6d41\u5f0f\u8f93\u51fa\uff0c\u6240\u4ee5\u6211\u4eec\u65e0\u9700\u518d\u63a5\u6536\u5176\u8fd4\u56de\u7684\u6570\u636e\uff0c\u76f4\u63a5\u8c03\u7528\u8fd9\u4e2a\u65b9\u6cd5\u5c31\u884c\uff1a<\/p>\n\n\n\n
        print(\"\\n--- \u751f\u6210\u7ed3\u679c ---\")\n    _ = model.generate(\n        **inputs,\n        max_new_tokens=max_tokens,\n        temperature=temperature,\n        top_p=top_p,\n        do_sample=True,\n        repetition_penalty=1.1,\n        streamer=streamer,\n    )\n    print(\"\\n--- \u751f\u6210\u5b8c\u6bd5 ---\")<\/code><\/pre>\n\n\n\n
    8. \u5b9e\u73b0main()\u65b9\u6cd5<\/h3><\/div>\n\n\n\n
    \u6700\u540e\u6211\u4eec\u6765\u5b9e\u73b0main()\u65b9\u6cd5\uff0c\u5c06\u521a\u521a\u51c6\u5907\u7684\u65b9\u6cd5\u90fd\u7528\u8d77\u6765\uff1a<\/p>\n\n\n\n
    \u8fd9\u4e2a\u5730\u65b9\u6709\u4e00\u70b9\u9700\u8981\u6ce8\u610f<\/strong>\uff0c\u5982\u679c\u8bfb\u8005\u662f\u5728Hugging Face\u4e0a\u9762\u901a\u8fc7git\u4e0b\u8f7d\u7684\u7b14\u8005\u5df2\u7ecf\u8bad\u7ec3\u597d\u7684\u6a21\u578b\uff0c\u8bf7\u628a\u4f60\u7684\u6a21\u578b\u6240\u5728\u7684\u76ee\u5f55\u8bbe\u7f6e\u597d\uff0c\u4e5f\u5c31\u662f_default_lora = “lora\/qwen3.5-9b-text-lora-checkpoint”\u9700\u8981\u6539\u4e3a_default_lora = “lora\/zjml\/Qwen3.5-9B-Fanren-LoRA”\uff0c\u8bf7\u8bfb\u8005\u901a\u8fc7\u5b9e\u9645\u60c5\u51b5\u8fdb\u884c\u4fee\u6539\uff0c\u5f53\u7136\u4e5f\u53ef\u4ee5\u901a\u8fc7\u547d\u4ee4\u4f20\u5165\uff1a<\/p>\n\n\n\n
    def main():\n    parser = argparse.ArgumentParser(description=\"qwopus3.5-9b-fanren-lora \u63a8\u7406\u811a\u672c\")\n    # \u667a\u80fd\u68c0\u6d4b LoRA \u8def\u5f84\uff1aHF clone \u5219\u5728\u5f53\u524d\u76ee\u5f55\uff0c\u672c\u5730\u5f00\u53d1\u5219\u5728 lora \u5b50\u76ee\u5f55\n    _default_lora = \".\"\n    if not Path(\"adapter_model.safetensors\").exists():\n        _default_lora = \"lora\/qwen3.5-9b-text-lora-checkpoint\"\n    parser.add_argument(\"--lora\", \"-l\", default=_default_lora,\n                        help=f\"LoRA \u9002\u914d\u5668\u8def\u5f84 (\u9ed8\u8ba4: {_default_lora})\")\n    parser.add_argument(\"--input\", \"-i\", default=None,\n                        help=\"\u524d\u6587\u5185\u5bb9\")\n    parser.add_argument(\"--max-tokens\", \"-n\", type=int, default=512,\n                        help=\"\u6700\u5927\u751f\u6210 token \u6570 (\u9ed8\u8ba4: 512)\")\n    parser.add_argument(\"--temperature\", type=float, default=0.8,\n                        help=\"\u6e29\u5ea6\u53c2\u6570\uff0c\u8d8a\u9ad8\u8d8a\u6709\u521b\u9020\u6027 (\u9ed8\u8ba4\u6309\u6848\u4f8b\u9884\u8bbe)\")\n    parser.add_argument(\"--top-p\", type=float, default=0.9,\n                        help=\"top_p \u91c7\u6837 (\u9ed8\u8ba4: 0.9)\")\n\n    args = parser.parse_args()\n\n    temp = args.temperature\n    tokens = args.max_tokens\n\n    print(f\"   \u6e29\u5ea6={temp}  tokens={tokens}\")\n\n    model, tokenizer = load_model(args.lora)\n    prompt = PROMPT_TEMPLATE.format(input_text=args.input)\n    generate(model, tokenizer, prompt, tokens, temp, args.top_p)<\/code><\/pre>\n\n\n\n
    \u6700\u540e\uff0c\u518d\u8bbe\u7f6e\u4e00\u4e2a\u7a0b\u5e8f\u5165\u53e3\uff1a<\/p>\n\n\n\n
    if __name__ == \"__main__\":\n    main()<\/code><\/pre>\n\n\n\n
    9. \u6267\u884c\u7eed\u5199<\/h3><\/div>\n\n\n\n
    \u7ec8\u4e8e\u5230\u4e86\u68c0\u9a8c\u6210\u679c\u7684\u65f6\u5019\u4e86\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u7b80\u5355\u7684\u547d\u4ee4\uff0c\u8ba9\u6a21\u578b\u4e3a\u6211\u4eec\u7684\u8f93\u5165\u505a\u4e00\u4e2a\u7eed\u5199\uff1a<\/p>\n\n\n\n
    python inference2.py -i \"\u97e9\u7acb\u72ec\u81ea\u884c\u8d70\u5728\u4e71\u661f\u6d77\u7684\u591c\u7a7a\u4e4b\u4e0b\u2014\u2014\"<\/code><\/pre>\n\n\n\n
    \u7b49\u5f85\u4e00\u6bb5\u65f6\u95f4\uff0c\u53ef\u4ee5\u770b\u5230\u6a21\u578b\u5df2\u7ecf\u6210\u529f\u5730\u8fdb\u884c\u6d41\u5f0f\u8f93\u51fa\u4e86\uff1a<\/p>\n\n\n\n
    \u5012\u8ba9\u539f\u672c\u4e00\u5fc3\u60f3\u8981\u8fdb\u5165\u91cc\u9762\u53d6\u5b9d\u7684\u97e9\u7acb\uff0c\u5fc3\u4e2d\u4e00\u52a8\uff0c\u9887\u6709\u5174\u8da3\u7684\u4ed4\u7ec6\u6253\u91cf\u8d77\u6b64\u5c71\u6765\u3002\n\n\n\u7b2c\u4e00\u5343\u4e00\u767e\u96f6\u4e03\u7ae0 \u5929\u90fd\u5c38\u706b\u4e0e\u5929\u7f61\u8840\u96f7\n\u3000\u3000\u6b64\u5c71\u867d\u7136\u9ad8\u5927\u5f02\u5e38\uff0c\u4f46\u660e\u663e\u662f\u88ab\u4eba\u7528\u83ab\u5927\u795e\u901a\u5f3a\u884c\u62ac\u5347\u5230\u534a\u7a7a\u4e2d\u4e00\u822c\uff0c\u5e95\u90e8\u4ecd\u6df1\u57cb\u5730\u4e0b\u6570\u767e\u4e08\u7684\u6837\u5b50\u3002\u5c71\u4f53\u8868\u9762\u95ea\u52a8\u7740\u6de1\u6de1\u7075\u5149\uff0c\u5e76\u6709\u4e00\u5c42\u51dd\u539a\u7684\u4e73\u767d\u8272\u51b0\u971c\u8986\u76d6\u5176\u4e0a\uff0c\u7adf\u5c06\u6574\u5ea7\u5de8\u5c71\u51b0\u5c01\u4e86\u8d77\u6765\u3002\n\u3000\u3000\u5982\u6b64\u505a\u7684\u597d\u5904\u81ea\u7136\u662f\u9632\u6b62\u91cc\u9762\u7684\u5b9d\u7269\u88ab\u4ed6\u4eba\u7aa5\u89c6\uff0c\u540c\u65f6\u4e5f\u53ef\u907f\u514d\u4e00\u4e9b\u6709\u5fc3\u4eba\u7684\u7834\u574f\u3002\u4e0d\u8fc7\u82e5\u662f\u56e0\u6b64\u771f\u80fd\u5c06\u90a3\u4e9b\u5b9d\u7269\u62a4\u4f4f\uff0c\u8fd9\u4e9b\u5316\u795e\u4fee\u58eb\u4e5f\u5c31\u6ca1\u6709\u5192\u5947\u9669\u8fdb\u5165\u5176\u4e2d\u7684\u610f\u601d\u4e86\u3002\n\u3000\u3000\u201c\u8fd9\u5c31\u662f\u5317\u6781\u5143\u5149\u7684\u5bd2\u529b\uff1f\u201d\u97e9\u7acb\u6b63\u60f3\u7684\u65f6\u5019\uff0c\u8033\u8fb9\u5374\u4f20\u6765\u4e86\u5927\u884d\u795e\u541b\u7684\u58f0\u97f3\uff1a\u201c\u4e0d\u9519\uff0c\u770b\u6765\u8001\u592b\u5f53\u521d\u6240\u8a00\u4e0d\u5047\u3002\u6b64\u5730\u7684\u786e\u6ca1\u6709\u4ec0\u4e48\u5317\u6781\u5143\u5149\u5b58\u5728\uff0c\u53ea\u662f\u4e0d\u77e5\u4e3a\u4f55\uff0c\u8fd9\u91cc\u5168\u90fd\u88ab\u4e00\u79cd\u6781\u5bd2\u4e4b\u529b\u5c01\u5370\u4f4f\u4e86\u3002\u201d\n\u3000\u3000\u201c\u662f\u5417\uff1f\u90a3\u665a\u8f88\u8fd9\u5c31\u89c1\u8bc6\u4e00\u4e0b\u4e86\u3002\u201d\u97e9\u7acb\u795e\u8272\u4e00\u52a8\uff0c\u5355\u624b\u4e00\u6390\u8bc0\uff0c\u987f\u65f6\u8eab\u4e0a\u9752\u5149\u4e00\u95ea\uff0c\u4e00\u5c42\u9752\u6fdb\u6fdb\u7684\u5149\u534e\u6d8c\u51fa\u3002\n\u3000\u3000 \u201c\u4e0d\uff01\u201d\n\u3000\u3000\u201c\u4ec0 \u4e48\uff1f\u201d\n\u3000\u3000\u97e9\u7acb\u7684\u8bdd\u97f3\u624d\u521a\u843d\uff0c\u5c31\u542c\u5f97\u5927\u884d\u795e\u541b\u4e00\u58f0\u60ca\u547c\u3002\u97e9\u7acb\u5fc3\u4e2d\u4e00\u51db\uff0c\u6025\u5fd9\u5b9a\u775b\u671b\u53bb\uff0c\u7ed3\u679c\u77b3\u5b54\u9aa4\u7136\u95f4\u4e00\u7f29\u3002\n\u3000\u3000\u53ea\u89c1\u5728\u4ed6\u8eab\u524d\u4e0d\u8fdc\u5904\uff0c\u4e00\u53ea\u9752\u8272\u5927\u624b\u51ed\u7a7a\u51fa\u73b0\uff0c\u95ea\u7535\u822c\u7684\u5411\u4e0b\u4e00\u6293\uff0c\u5c31\u5c06\u4e00\u53ea\u767d\u8272\u5c0f\u9f0e\u4e00\u628a\u635e\u5728\u4e86\u624b\u4e2d\u3002\n\u3000\u3000\u800c\u5c0f\u9f0e\u56db\u5468\u7684\u7d2b\u8272\u706b\u7130\uff0c\u5728\u9752\u8272\u5927\u624b\u65b9\u4e00\u63a5\u89e6\u7684\u77ac\u95f4\uff0c\u7acb\u523b\u88ab\u51bb\u7ed3\u6210\u4e00\u5757\u7d2b\u8272\u51b0\u5757\u3002\u6574\u53ea\u5927\u624b\u4e5f\u77ac\u95f4\u53d8\u5f97\u6676\u83b9\u95ea\u70c1\uff0c\u5bd2\u6c14\u903c\u4eba\u3002\n\u3000\u3000\u97e9\u7acb\u5fc3\u4e2d\u4e00\u60ca\uff0c\u4e0d\u53ca\u591a\u60f3\u7684\u53e3\u4e2d\u4e00\u58f0\u5927\u559d\u3002\n\u3000\u3000\u4e00\u80a1\u9ed1\u6c14\u4ece\u4e39\u7530\u5904\u98de\u7a9c\u800c\u51fa\uff0c\u968f\u5373\u5316\u4e3a\u4e00\u6761\u9ed1\u8272\u5c0f\u86df\u51b2\u5411\u4e86\u90a3\u53ea\u9752\u8272\u5927\u624b\u3002\u7ed3\u679c\u5728\u4e00\u63a5\u89e6\u4e0b\uff0c\u7acb\u523b\u88ab\u51bb\u6210\u4e86\u53e6\u4e00\u5757\u5de8\u5927\u51b0\u5757\u3002\n\u3000\u3000\u97e9\u7acb\u8138\u8272\u6709\u4e9b\u96be\u770b\uff0c\u76ee\u4e2d\u95ea\u8fc7\u4e00\u4e1d\u8be7\u5f02\u4e4b\u8272\u3002\n\u3000\u3000\u4ed6\u81ea\u4ece\u4fee\u6210\u4e86\u715e\u4e39\u540e\uff0c\u53ef\u4ece\u672a\u51fa\u73b0\u8fc7\u8fd9\u79cd\u60c5\u5f62\u3002\u96be\u9053\u8fd9\u5317\u6781\u5143\u5149\u5177\u6709\u4ec0\u4e48\u7279\u6b8a\u795e\u901a\u4e0d\u6210\uff1f\n\u3000\u3000\u5fc3\u5ff5\u6025\u8f6c\u95f4\uff0c\u97e9\u7acb\u795e\u8bc6\u4e00\u52a8\u4e0b\uff0c\u5468\u8eab\u6d6e\u73b0\u7684\u9752\u8272\u7075\u7eb9\uff0c\u540c\u65f6\u653e\u51fa\u523a\u76ee\u7684\u7075\u8292\u3002\n\u3000\u3000\u5404\u8272\u7b26\u6587\u4ea4\u7ec7\u4e4b\u4e0b\uff0c\u97e9\u7acb\u8eab\u5f62\u7565\u4e00\u6a21\u7cca\uff0c\u84e6\u7136\u4ece\u539f\u5730\n\n--- \u751f\u6210\u5b8c\u6bd5 ---<\/code><\/pre>\n\n\n\n
    \u53ef\u4ee5\u770b\u5230\uff0c\u4f5c\u4e3a\u7eed\u5199\u7684\u8bdd\uff0c\u6a21\u578b\u8fd8\u662f\u6bd4\u8f83\u6210\u529f\u7684\uff0c\u81f3\u5c11\u8bed\u53e5\u6ca1\u6709\u4ec0\u4e48\u95ee\u9898\u3002\u5e76\u4e14\u4e5f\u80fd\u591f\u6b63\u786e\u5730\u4f7f\u7528\u51e1\u4eba\u4fee\u4ed9\u4f20\u4e2d\u7684\u89d2\u8272\u540d\u79f0\uff0c\u5176\u5199\u4f5c\u98ce\u683c\u4e5f\u4e0e\u51e1\u4eba\u4fee\u4ed9\u4f20\u975e\u5e38\u76f8\u4f3c\uff0c\u8fd9\u5c31\u662f\u5fae\u8c03\u6240\u5b9e\u73b0\u7684\u6548\u679c\u3002<\/p>\n\n\n\n
    \u7ed3\u8bed<\/h2><\/div>\n\n\n\n
    \u5230\u73b0\u5728\uff0c\u76f8\u4fe1\u5404\u4f4d\u8bfb\u8005\u5df2\u7ecf\u80fd\u591f\u72ec\u7acb\u5b9e\u73b0\u6a21\u578b\u7684\u5fae\u8c03\u4e86\u3002\u6a21\u578b\u7684\u5fae\u8c03\u5728\u5b9e\u9645\u4f7f\u7528\u4e2d\u6709\u7740\u5f88\u5927\u7684\u6f5c\u529b\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u5c0f\u6a21\u578b\uff0c\u7ecf\u8fc7\u5fae\u8c03\u4e4b\u540e\uff0c\u4e5f\u80fd\u5728\u5176\u7279\u5b9a\u7684\u9886\u57df\u5c55\u73b0\u51fa\u5f88\u5f3a\u7684\u5b9e\u9645\u6548\u679c\uff0c\u5e76\u4e14\u80fd\u5927\u5927\u5730\u51cf\u5c11\u7b97\u529b\u4e0e\u663e\u5b58\u7684\u4f7f\u7528\u3002\u4e0e\u63d0\u793a\u8bcd\u5de5\u7a0b\u4e0d\u540c\uff0c\u6a21\u578b\u5fae\u8c03\u5728\u66f4\u6df1\u5c42\u6b21\u4e2d\u5bf9\u6a21\u578b\u7684\u53c2\u6570\u505a\u51fa\u4e86\u6539\u53d8\uff0c\u5e76\u4e14\u76f8\u6bd4\u63d0\u793a\u8bcd\u5de5\u7a0b\uff0c\u6211\u4eec\u65e0\u9700\u62c5\u5fc3\u6a21\u578b\u7531\u4e8e\u957f\u4e0a\u4e0b\u6587\u4e2d\u7684\u6ce8\u610f\u529b\u673a\u5236\u6216\u8005\u9057\u5fd8\u673a\u5236\uff0c\u5bfc\u81f4\u6a21\u578b\u7684\u8f93\u51fa\u504f\u79bb\u6211\u4eec\u7684\u9884\u671f\uff0c\u5fae\u8c03\u7684\u5c0f\u6a21\u578b\u5728\u6d88\u8d39\u7ea7\u7684\u5bb6\u7528\u8ba1\u7b97\u673a\u4e2d\u4e5f\u80fd\u6709\u8f83\u597d\u7684\u8f93\u51fa\u8868\u73b0\u3002<\/p>\n\n\n\n
    \u5e76\u4e14\u7ecf\u8fc7LoRA \u548cUnsloth \u9879\u76ee\u7684\u9876\u5c16\u5de5\u7a0b\u5e08\u7684\u8c03\u6559\uff0c\u5728\u5bb6\u7528\u7ea7\u663e\u5361\u4e0a\u4e5f\u80fd\u5bf9\u8f83\u5927\u53c2\u6570\u7684\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u65e0\u9700\u6211\u4eec\u72ec\u7acb\u91cd\u65b0\u8bad\u7ec3\u4e00\u4e2a\u65b0\u7684\u6a21\u578b\u3002\u5bf9\u4e8e\u5bb6\u7528\u6d88\u8d39\u7ea7\u663e\u5361\uff0c\u6211\u4eec\u662f\u5b8c\u5168\u4e0d\u53ef\u80fd\u8fdb\u884c\u5168\u91cf\u5fae\u8c03\u6216\u8005\u72ec\u7acb\u8bad\u7ec3\u4e00\u4e2a9B\u53c2\u6570\u7684\u6a21\u578b\uff0c\u4f46\u6211\u4eec\u5374\u53ef\u4ee5\u5bf9\u5176\u8fdb\u884c\u5fae\u8c03\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u6211\u4eec\u65e0\u9700\u6559\u4f1a\u6a21\u578b\u8bf4\u8bdd\u7684\u80fd\u529b\uff0c\u4f46\u6211\u4eec\u5374\u53ef\u4ee5\u5bf9\u5176\u8bf4\u8bdd\u7684\u98ce\u683c\u505a\u51fa\u4e00\u4e9b\u7ec6\u5c0f\u7684\u6539\u53d8\u3002\u53e6\u5916\uff0c\u6a21\u578b\u5fae\u8c03\u53ea\u9700\u4f7f\u7528\u8f83\u5c0f\u7684\u6587\u672c\u6837\u672c\u91cf\uff0c\u666e\u901a\u7528\u6237\u8f83\u96be\u6536\u96c6\u51e0\u5341G\u751a\u81f3\u51e0TB\u7684\u4f18\u8d28\u6587\u672c\u6570\u636e\u7528\u4e8e\u6559\u4f1a\u6a21\u578b\u8bf4\u8bdd\uff0c\u800c\u6a21\u578b\u5fae\u8c03\u5f80\u5f80\u53ea\u9700\u8981\u4f7f\u7528\u51e0\u767e\u751a\u81f3\u51e0\u5341MB\u7684\u6587\u672c\u4fbf\u53ef\u4ee5\u8fdb\u884c\u5fae\u8c03\uff0c\u5176\u96be\u5ea6\u4e0b\u964d\u4e86\u6570\u4e2a\u6570\u91cf\u7ea7\u3002<\/p>\n\n\n\n
    \u4ee5\u4e0a\u5c31\u662f\u672c\u6587\u7684\u5168\u90e8\u5185\u5bb9\u4e86\uff0c\u5199\u4f5c\u4e0d\u6613\uff0c\u8fd8\u671b\u5927\u5bb6\u591a\u591a\u652f\u6301\uff0c\u5e76\u4e14\u53ef\u4ee5\u5728\u8bc4\u8bba\u533a\u8fdb\u884c\u8ba8\u8bba\uff0c\u611f\u5174\u8da3\u7684\u670b\u53cb\u53ef\u4ee5\u901a\u8fc7hnoycy@mymailbox.cn \u4e0e\u7b14\u8005\u53d6\u5f97\u8054\u7cfb\uff0c\u5171\u540c\u8fdb\u6b65\uff01<\/p>\n","protected":false},"excerpt":{"rendered":"
    \u4ec0\u4e48\u662fLoRA\u5fae\u8c03 \u4e0eLoRA\u5fae\u8c03\u5bf9\u5e94\u7684\u662f\u5168\u91cf\u5fae\u8c03\uff0c\u5373\u5bf9\u6574\u4e2a\u6a21\u578b\u6240\u6709\u7684\u53c2\u6570\u6839\u636e\u8f93\u5165\u7684\u6587\u672c\u8fdb\u884c\u5fae\u8c03\u66f4\u65b0\u3002\u8fd9\u4e2a\u6a21\u578b […]<\/p>\n","protected":false},"author":1,"featured_media":2331,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_uag_custom_page_level_css":"","_monsterinsights_skip_tracking":false,"footnotes":""},"categories":[10],"tags":[24],"class_list":["post-2331","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10","tag-24"],"uagb_featured_image_src":{"full":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",800,600,false],"thumbnail":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",150,150,true],"medium":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",300,300,false],"medium_large":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",768,0,false],"large":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",1024,1024,false],"1536x1536":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",1536,1536,false],"2048x2048":["https:\/\/pic.zjml.life\/api\/image\/1785054361862b28c29?quality=high",2048,2048,false]},"uagb_author_info":{"display_name":"zjml","author_link":"https:\/\/blog.zjml.life\/?author=1"},"uagb_comment_info":0,"uagb_excerpt":"\u4ec0\u4e48\u662fLoRA\u5fae\u8c03 \u4e0eLoRA\u5fae\u8c03\u5bf9\u5e94\u7684\u662f\u5168\u91cf\u5fae\u8c03\uff0c\u5373\u5bf9\u6574\u4e2a\u6a21\u578b\u6240\u6709\u7684\u53c2\u6570\u6839\u636e\u8f93\u5165\u7684\u6587\u672c\u8fdb\u884c\u5fae\u8c03\u66f4\u65b0\u3002\u8fd9\u4e2a\u6a21\u578b…","_links":{"self":[{"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/posts\/2331","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2331"}],"version-history":[{"count":28,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/posts\/2331\/revisions"}],"predecessor-version":[{"id":2368,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/posts\/2331\/revisions\/2368"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=\/wp\/v2\/posts\/2331"}],"wp:attachment":[{"href":"https:\/\/blog.zjml.life\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2331"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2331"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.zjml.life\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2331"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}