{"id":24120,"date":"2025-01-21T09:51:35","date_gmt":"2025-01-21T01:51:35","guid":{"rendered":"http:\/\/139.9.1.231\/?p=24120"},"modified":"2025-01-21T09:51:36","modified_gmt":"2025-01-21T01:51:36","slug":"qwen2-lora","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2025\/01\/21\/qwen2-lora\/","title":{"rendered":"Qwen2 Lora LLM\u5fae\u8c03\u8bad\u7ec3\u6559\u7a0b"},"content":{"rendered":"\n

Github\uff1a<\/strong>https:\/\/github.com\/datawhalechina\/self-llm\/tree\/master<\/em><\/strong><\/a><\/p>\n\n\n\n\n\n

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

\u2003\u672c\u9879\u76ee\u662f\u4e00\u4e2a\u56f4\u7ed5\u5f00\u6e90\u5927\u6a21\u578b\u3001\u9488\u5bf9\u56fd\u5185\u521d\u5b66\u8005\u3001\u57fa\u4e8e Linux \u5e73\u53f0\u7684\u4e2d\u56fd\u5b9d\u5b9d\u4e13\u5c5e\u5927\u6a21\u578b\u6559\u7a0b\uff0c\u9488\u5bf9\u5404\u7c7b\u5f00\u6e90\u5927\u6a21\u578b\u63d0\u4f9b\u5305\u62ec\u73af\u5883\u914d\u7f6e\u3001\u672c\u5730\u90e8\u7f72\u3001\u9ad8\u6548\u5fae\u8c03\u7b49\u6280\u80fd\u5728\u5185\u7684\u5168\u6d41\u7a0b\u6307\u5bfc\uff0c\u7b80\u5316\u5f00\u6e90\u5927\u6a21\u578b\u7684\u90e8\u7f72\u3001\u4f7f\u7528\u548c\u5e94\u7528\u6d41\u7a0b\uff0c\u8ba9\u66f4\u591a\u7684\u666e\u901a\u5b66\u751f\u3001\u7814\u7a76\u8005\u66f4\u597d\u5730\u4f7f\u7528\u5f00\u6e90\u5927\u6a21\u578b\uff0c\u5e2e\u52a9\u5f00\u6e90\u3001\u81ea\u7531\u7684\u5927\u6a21\u578b\u66f4\u5feb\u878d\u5165\u5230\u666e\u901a\u5b66\u4e60\u8005\u7684\u751f\u6d3b\u4e2d\u3002<\/p>\n\n\n\n

\u2003\u2003\u672c\u9879\u76ee\u7684\u4e3b\u8981\u5185\u5bb9\u5305\u62ec\uff1a<\/p>\n\n\n\n

  1. \u57fa\u4e8e Linux \u5e73\u53f0\u7684\u5f00\u6e90 LLM \u73af\u5883\u914d\u7f6e\u6307\u5357\uff0c\u9488\u5bf9\u4e0d\u540c\u6a21\u578b\u8981\u6c42\u63d0\u4f9b\u4e0d\u540c\u7684\u8be6\u7ec6\u73af\u5883\u914d\u7f6e\u6b65\u9aa4\uff1b<\/li>
  2. \u9488\u5bf9\u56fd\u5185\u5916\u4e3b\u6d41\u5f00\u6e90 LLM \u7684\u90e8\u7f72\u4f7f\u7528\u6559\u7a0b\uff0c\u5305\u62ec LLaMA\u3001ChatGLM\u3001InternLM \u7b49\uff1b<\/li>
  3. \u5f00\u6e90 LLM \u7684\u90e8\u7f72\u5e94\u7528\u6307\u5bfc\uff0c\u5305\u62ec\u547d\u4ee4\u884c\u8c03\u7528\u3001\u5728\u7ebf Demo \u90e8\u7f72\u3001LangChain \u6846\u67b6\u96c6\u6210\u7b49\uff1b<\/li>
  4. \u5f00\u6e90 LLM \u7684\u5168\u91cf\u5fae\u8c03\u3001\u9ad8\u6548\u5fae\u8c03\u65b9\u6cd5\uff0c\u5305\u62ec\u5206\u5e03\u5f0f\u5168\u91cf\u5fae\u8c03\u3001LoRA\u3001ptuning \u7b49\u3002<\/li><\/ol>\n\n\n\n

    \u73af\u5883\u914d\u7f6e<\/a><\/h2>\n\n\n\n

    \u5728\u5b8c\u6210\u57fa\u672c\u73af\u5883\u914d\u7f6e\u548c\u672c\u5730\u6a21\u578b\u90e8\u7f72\u7684\u60c5\u51b5\u4e0b\uff0c\u4f60\u8fd8\u9700\u8981\u5b89\u88c5\u4e00\u4e9b\u7b2c\u4e09\u65b9\u5e93\uff0c\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/p>\n\n\n\n

    python -m pip install --upgrade pip\n# \u66f4\u6362 pypi \u6e90\u52a0\u901f\u5e93\u7684\u5b89\u88c5\npip config set global.index-url https:\/\/pypi.tuna.tsinghua.edu.cn\/simple\n\npip install modelscope==1.18.0\npip install transformers==4.44.2\npip install streamlit==1.24.0\npip install sentencepiece==0.2.0\npip install accelerate==0.34.2\npip install datasets==2.20.0\npip install peft==0.11.1<\/pre>\n\n\n\n
    \u6a21\u578b\u4e0b\u8f7d<\/a><\/h2>\n\n\n\n
    \u4f7f\u7528 modelscope \u4e2d\u7684 snapshot_download \u51fd\u6570\u4e0b\u8f7d\u6a21\u578b\uff0c\u7b2c\u4e00\u4e2a\u53c2\u6570\u4e3a\u6a21\u578b\u540d\u79f0\uff0c\u53c2\u6570 cache_dir \u4e3a\u6a21\u578b\u7684\u4e0b\u8f7d\u8def\u5f84\u3002<\/p>\n\n\n\n
    \u5728 \/root\/autodl-tmp \u8def\u5f84\u4e0b\u65b0\u5efa model_download.py \u6587\u4ef6\u5e76\u5728\u5176\u4e2d\u8f93\u5165\u4ee5\u4e0b\u5185\u5bb9\uff0c\u7c98\u8d34\u4ee3\u7801\u540e\u8bf7\u53ca\u65f6\u4fdd\u5b58\u6587\u4ef6\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002\u5e76\u8fd0\u884c python \/root\/autodl-tmp\/model_download.py<\/code> \u6267\u884c\u4e0b\u8f7d\uff0c\u6a21\u578b\u5927\u5c0f\u4e3a 15GB\uff0c\u4e0b\u8f7d\u6a21\u578b\u5927\u6982\u9700\u8981 5 \u5206\u949f\u3002<\/p>\n\n\n\n
    import torch\nfrom modelscope import snapshot_download, AutoModel, AutoTokenizer\nimport os\nmodel_dir = snapshot_download('qwen\/Qwen2.5-7B-Instruct', cache_dir='\/root\/autodl-tmp', revision='master')<\/pre>\n\n\n\n
    \u6307\u4ee4\u96c6\u6784\u5efa<\/a><\/h2>\n\n\n\n
    LLM \u7684\u5fae\u8c03\u4e00\u822c\u6307\u6307\u4ee4\u5fae\u8c03\u8fc7\u7a0b\u3002\u6240\u8c13\u6307\u4ee4\u5fae\u8c03\uff0c\u662f\u8bf4\u6211\u4eec\u4f7f\u7528\u7684\u5fae\u8c03\u6570\u636e\u5f62\u5982\uff1a<\/p>\n\n\n\n
    {\n  \"instruction\": \"\u56de\u7b54\u4ee5\u4e0b\u7528\u6237\u95ee\u9898\uff0c\u4ec5\u8f93\u51fa\u7b54\u6848\u3002\",\n  \"input\": \"1+1\u7b49\u4e8e\u51e0?\",\n  \"output\": \"2\"\n}<\/pre>\n\n\n\n
    \u5176\u4e2d\uff0cinstruction<\/code> \u662f\u7528\u6237\u6307\u4ee4\uff0c\u544a\u77e5\u6a21\u578b\u5176\u9700\u8981\u5b8c\u6210\u7684\u4efb\u52a1\uff1binput<\/code> \u662f\u7528\u6237\u8f93\u5165\uff0c\u662f\u5b8c\u6210\u7528\u6237\u6307\u4ee4\u6240\u5fc5\u987b\u7684\u8f93\u5165\u5185\u5bb9\uff1boutput<\/code> \u662f\u6a21\u578b\u5e94\u8be5\u7ed9\u51fa\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n
    \u5373\u6211\u4eec\u7684\u6838\u5fc3\u8bad\u7ec3\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u5177\u6709\u7406\u89e3\u5e76\u9075\u5faa\u7528\u6237\u6307\u4ee4\u7684\u80fd\u529b\u3002\u56e0\u6b64\uff0c\u5728\u6307\u4ee4\u96c6\u6784\u5efa\u65f6\uff0c\u6211\u4eec\u5e94\u9488\u5bf9\u6211\u4eec\u7684\u76ee\u6807\u4efb\u52a1\uff0c\u9488\u5bf9\u6027\u6784\u5efa\u4efb\u52a1\u6307\u4ee4\u96c6\u3002\u4f8b\u5982\uff0c\u5728\u672c\u8282\u6211\u4eec\u4f7f\u7528\u7531\u7b14\u8005\u5408\u4f5c\u5f00\u6e90\u7684 Chat-\u7504\u5b1b<\/a> \u9879\u76ee\u4f5c\u4e3a\u793a\u4f8b\uff0c\u6211\u4eec\u7684\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u80fd\u591f\u6a21\u62df\u7504\u5b1b\u5bf9\u8bdd\u98ce\u683c\u7684\u4e2a\u6027\u5316 LLM\uff0c\u56e0\u6b64\u6211\u4eec\u6784\u9020\u7684\u6307\u4ee4\u5f62\u5982\uff1a<\/p>\n\n\n\n
    {\n  \"instruction\": \"\u4f60\u662f\u8c01\uff1f\",\n  \"input\": \"\",\n  \"output\": \"\u5bb6\u7236\u662f\u5927\u7406\u5bfa\u5c11\u537f\u7504\u8fdc\u9053\u3002\"\n}<\/pre>\n\n\n\n
    \u6211\u4eec\u6240\u6784\u9020\u7684\u5168\u90e8\u6307\u4ee4\u6570\u636e\u96c6\u5728\u6839\u76ee\u5f55\u4e0b\u3002<\/p>\n\n\n\n
    \u6570\u636e\u683c\u5f0f\u5316<\/h2>\n\n\n\n
    <\/a><\/p>\n\n\n\n
    Lora<\/code> \u8bad\u7ec3\u7684\u6570\u636e\u662f\u9700\u8981\u7ecf\u8fc7\u683c\u5f0f\u5316\u3001\u7f16\u7801\u4e4b\u540e\u518d\u8f93\u5165\u7ed9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\u7684\uff0c\u5982\u679c\u662f\u719f\u6089 Pytorch<\/code> \u6a21\u578b\u8bad\u7ec3\u6d41\u7a0b\u7684\u540c\u5b66\u4f1a\u77e5\u9053\uff0c\u6211\u4eec\u4e00\u822c\u9700\u8981\u5c06\u8f93\u5165\u6587\u672c\u7f16\u7801\u4e3a input_ids\uff0c\u5c06\u8f93\u51fa\u6587\u672c\u7f16\u7801\u4e3a labels<\/code>\uff0c\u7f16\u7801\u4e4b\u540e\u7684\u7ed3\u679c\u90fd\u662f\u591a\u7ef4\u7684\u5411\u91cf\u3002\u6211\u4eec\u9996\u5148\u5b9a\u4e49\u4e00\u4e2a\u9884\u5904\u7406\u51fd\u6570\uff0c\u8fd9\u4e2a\u51fd\u6570\u7528\u4e8e\u5bf9\u6bcf\u4e00\u4e2a\u6837\u672c\uff0c\u7f16\u7801\u5176\u8f93\u5165\u3001\u8f93\u51fa\u6587\u672c\u5e76\u8fd4\u56de\u4e00\u4e2a\u7f16\u7801\u540e\u7684\u5b57\u5178\uff1a<\/p>\n\n\n\n
    def process_func(example):\n    MAX_LENGTH = 384    # Llama\u5206\u8bcd\u5668\u4f1a\u5c06\u4e00\u4e2a\u4e2d\u6587\u5b57\u5207\u5206\u4e3a\u591a\u4e2atoken\uff0c\u56e0\u6b64\u9700\u8981\u653e\u5f00\u4e00\u4e9b\u6700\u5927\u957f\u5ea6\uff0c\u4fdd\u8bc1\u6570\u636e\u7684\u5b8c\u6574\u6027\n    input_ids, attention_mask, labels = [], [], []\n    instruction = tokenizer(f\"<|im_start|>system\\n\u73b0\u5728\u4f60\u8981\u626e\u6f14\u7687\u5e1d\u8eab\u8fb9\u7684\u5973\u4eba--\u7504\u5b1b<|im_end|>\\n<|im_start|>user\\n{example['instruction'] + example['input']}<|im_end|>\\n<|im_start|>assistant\\n\", add_special_tokens=False)  # add_special_tokens \u4e0d\u5728\u5f00\u5934\u52a0 special_tokens\n    response = tokenizer(f\"{example['output']}\", add_special_tokens=False)\n    input_ids = instruction[\"input_ids\"] + response[\"input_ids\"] + [tokenizer.pad_token_id]\n    attention_mask = instruction[\"attention_mask\"] + response[\"attention_mask\"] + [1]  # \u56e0\u4e3aeos token\u54b1\u4eec\u4e5f\u662f\u8981\u5173\u6ce8\u7684\u6240\u4ee5 \u8865\u5145\u4e3a1\n    labels = [-100] * len(instruction[\"input_ids\"]) + response[\"input_ids\"] + [tokenizer.pad_token_id]\n    if len(input_ids) > MAX_LENGTH:  # \u505a\u4e00\u4e2a\u622a\u65ad\n        input_ids = input_ids[:MAX_LENGTH]\n        attention_mask = attention_mask[:MAX_LENGTH]\n        labels = labels[:MAX_LENGTH]\n    return {\n        \"input_ids\": input_ids,\n        \"attention_mask\": attention_mask,\n        \"labels\": labels\n    }<\/pre>\n\n\n\n
    Qwen2<\/code> \u91c7\u7528\u7684 Prompt Template<\/code>\u683c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n
    <|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n\u4f60\u662f\u8c01\uff1f<|im_end|>\n<|im_start|>assistant\n\u6211\u662f\u4e00\u4e2a\u6709\u7528\u7684\u52a9\u624b\u3002<|im_end|><\/code><\/pre>\n\n\n\n
    \u52a0\u8f7d tokenizer \u548c\u534a\u7cbe\u5ea6\u6a21\u578b<\/a><\/h2>\n\n\n\n
    \u6a21\u578b\u4ee5\u534a\u7cbe\u5ea6\u5f62\u5f0f\u52a0\u8f7d\uff0c\u5982\u679c\u4f60\u7684\u663e\u5361\u6bd4\u8f83\u65b0\u7684\u8bdd\uff0c\u53ef\u4ee5\u7528 torch.bfolat<\/code>\u5f62\u5f0f\u52a0\u8f7d\u3002\u5bf9\u4e8e\u81ea\u5b9a\u4e49\u7684\u6a21\u578b\u4e00\u5b9a\u8981\u6307\u5b9a trust_remote_code<\/code>\u53c2\u6570\u4e3a True<\/code>\u3002<\/p>\n\n\n\n
    tokenizer = AutoTokenizer.from_pretrained('\/root\/autodl-tmp\/qwen\/Qwen2.5-7B-Instruct\/', use_fast=False, trust_remote_code=True)\n\nmodel = AutoModelForCausalLM.from_pretrained('\/root\/autodl-tmp\/qwen\/Qwen2.5-7B-Instruct\/', device_map=\"auto\",torch_dtype=torch.bfloat16)<\/pre>\n\n\n\n
    \u5b9a\u4e49 LoraConfig<\/a><\/h2>\n\n\n\n
    LoraConfig<\/code>\u8fd9\u4e2a\u7c7b\u4e2d\u53ef\u4ee5\u8bbe\u7f6e\u5f88\u591a\u53c2\u6570\uff0c\u4f46\u4e3b\u8981\u7684\u53c2\u6570\u6ca1\u591a\u5c11\uff0c\u7b80\u5355\u8bb2\u4e00\u8bb2\uff0c\u611f\u5174\u8da3\u7684\u540c\u5b66\u53ef\u4ee5\u76f4\u63a5\u770b\u6e90\u7801\u3002<\/p>\n\n\n\n