{"id":1250,"date":"2026-03-30T06:38:09","date_gmt":"2026-03-30T14:38:09","guid":{"rendered":"https:\/\/51ai.website\/?p=1250"},"modified":"2026-03-30T06:38:12","modified_gmt":"2026-03-30T14:38:12","slug":"llm-model-deploy-fine-turning","status":"publish","type":"post","link":"https:\/\/51ai.website\/en\/blog\/llm-model-deploy-fine-turning\/","title":{"rendered":"LLM\u6a21\u578b\u90e8\u7f72\u4e0e\u5fae\u8c03"},"content":{"rendered":"

\u5df2\u7ecf\u4e86\u89e3\u4e86\u5927\u6a21\u578b\u7684\u57fa\u7840\u6982\u5ff5\u3001RAG\u3001Agent\u3001Transformer\u67b6\u6784\u7b49\u6838\u5fc3\u77e5\u8bc6\u3002\u4eca\u5929\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u9002\u914d\u5230\u81ea\u5df1\u7684\u4e1a\u52a1\u573a\u666f\uff08\u5fae\u8c03\uff09<\/strong>\uff0c\u5e76\u5c06\u5176\u7a33\u5b9a\u3001\u9ad8\u6548\u5730\u90e8\u7f72\u5230\u751f\u4ea7\u73af\u5883\uff08\u90e8\u7f72\uff09<\/strong>The<\/p>\n\n\n\n

\u5fae\u8c03\u8ba9\u6a21\u578b\u66f4\u61c2\u4f60\u7684\u4e1a\u52a1\uff0c\u90e8\u7f72\u8ba9\u6a21\u578b\u771f\u6b63\u670d\u52a1\u4e8e\u7528\u6237\u3002\u4e24\u8005\u7ed3\u5408\uff0c\u662fAI\u5e94\u7528\u843d\u5730\u7684\u6700\u540e\u4e00\u516c\u91cc\u3002<\/p>\n\n\n\n

\u4e00\u3001\u4e3a\u4ec0\u4e48\u9700\u8981\u5fae\u8c03\uff1f<\/strong><\/h2>\n\n\n\n

\u9884\u8bad\u7ec3\u5927\u6a21\u578b\uff08\u5982GPT-4\u3001LLaMA\uff09\u5df2\u7ecf\u5177\u5907\u901a\u7528\u80fd\u529b\uff0c\u4f46\u5728\u7279\u5b9a\u573a\u666f\u4e0b\u53ef\u80fd\u8868\u73b0\u4e0d\u4f73\uff1a<\/p>\n\n\n\n

\u25cf \u98ce\u683c\u4e0d\u7b26<\/strong>\uff1a\u56de\u7b54\u4e0d\u591f\u4e13\u4e1a\u3001\u8bed\u6c14\u4e0d\u5339\u914d\u3002<\/p>\n\n\n\n

\u25cf \u77e5\u8bc6\u7f3a\u5931<\/strong>\uff1a\u4e0d\u4e86\u89e3\u4f60\u7684\u4ea7\u54c1\u3001\u516c\u53f8\u5185\u90e8\u672f\u8bed\u3002<\/p>\n\n\n\n

\u25cf \u7ed3\u6784\u8981\u6c42<\/strong>\uff1a\u9700\u8981\u8f93\u51fa\u7279\u5b9a\u683c\u5f0f\uff08JSON\u3001\u8868\u683c\uff09\u3002<\/p>\n\n\n\n

\u25cf \u6210\u672c\u8003\u91cf<\/strong>\uff1a\u9891\u7e41\u8c03\u7528API\u6210\u672c\u9ad8\uff0c\u5fae\u8c03\u5c0f\u6a21\u578b\u66f4\u5212\u7b97\u3002<\/p>\n\n\n\n

1.1 \u5fae\u8c03 vs \u63d0\u793a\u5de5\u7a0b vs RAG<\/strong><\/h3>\n\n\n\n
\u65b9\u6cd5<\/strong><\/td>\u9002\u7528\u573a\u666f<\/strong><\/td>\u4f18\u70b9<\/strong><\/td>\u7f3a\u70b9<\/strong><\/td><\/tr>
\u63d0\u793a\u5de5\u7a0b<\/strong><\/td>\u4e34\u65f6\u3001\u7b80\u5355\u4efb\u52a1<\/td>\u65e0\u9700\u8bad\u7ec3\uff0c\u5373\u65f6\u751f\u6548<\/td>\u590d\u6742\u4efb\u52a1\u4e0d\u7a33\u5b9a\uff0cToken\u6d88\u8017\u5927<\/td><\/tr>
RAG<\/strong><\/td>\u77e5\u8bc6\u95ee\u7b54\u3001\u79c1\u6709\u6570\u636e<\/td>\u5b9e\u65f6\u66f4\u65b0\uff0c\u53ef\u89e3\u91ca<\/td>\u4f9d\u8d56\u68c0\u7d22\u8d28\u91cf\uff0c\u5ef6\u8fdf\u7a0d\u9ad8<\/td><\/tr>
\u5fae\u8c03<\/strong><\/td>\u98ce\u683c\u3001\u683c\u5f0f\u3001\u7279\u5b9a\u9886\u57df<\/td>\u6df1\u5ea6\u5b9a\u5236\uff0c\u6027\u80fd\u7a33\u5b9a<\/td>\u9700\u8981\u8bad\u7ec3\u6570\u636e\uff0c\u6210\u672c\u8f83\u9ad8<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\u4e09\u8005\u5e38\u7ed3\u5408\u4f7f\u7528\uff1a\u5fae\u8c03\u8ba9\u6a21\u578b\u201c\u5b66\u4f1a\u201d\u4f60\u7684\u9886\u57df\u77e5\u8bc6\uff0cRAG\u63d0\u4f9b\u5b9e\u65f6\u4fe1\u606f\uff0c\u63d0\u793a\u5de5\u7a0b\u4f18\u5316\u4ea4\u4e92\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u5fae\u8c03\u7684\u6838\u5fc3\u6d41\u7a0b<\/strong><\/h2>\n\n\n\n

\u5fae\u8c03\u662f\u5728\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u57fa\u7840\u4e0a\uff0c\u7528\u5c11\u91cf\u4efb\u52a1\u76f8\u5173\u6570\u636e\u7ee7\u7eed\u8bad\u7ec3\uff0c\u8c03\u6574\u6a21\u578b\u53c2\u6570\u3002<\/p>\n\n\n\n

2.1 \u5fae\u8c03\u6b65\u9aa4<\/strong><\/h3>\n\n\n\n

1. \u6570\u636e\u51c6\u5907<\/strong>\uff1a\u6536\u96c6\u3001\u6e05\u6d17\u3001\u683c\u5f0f\u5316\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n\n\n\n

2. \u9009\u62e9\u57fa\u5ea7\u6a21\u578b<\/strong>\uff1a\u6839\u636e\u9700\u6c42\u9009\u62e9\u5408\u9002\u6a21\u578b\uff08\u5982LLaMA\u3001ChatGLM\u3001Qwen\uff09\u3002<\/p>\n\n\n\n

3. \u5fae\u8c03\u65b9\u6cd5<\/strong>\uff1a\u5168\u91cf\u5fae\u8c03\u6216\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\uff08PEFT\uff09\u3002<\/p>\n\n\n\n

4. \u8bad\u7ec3<\/strong>\uff1a\u8bbe\u7f6e\u8d85\u53c2\u6570\uff0c\u542f\u52a8\u8bad\u7ec3\u3002<\/p>\n\n\n\n

5. \u8bc4\u4f30\u4e0e\u8fed\u4ee3<\/strong>\uff1a\u7528\u9a8c\u8bc1\u96c6\u8bc4\u4f30\u6548\u679c\uff0c\u8c03\u6574\u6570\u636e\u6216\u53c2\u6570\u3002<\/p>\n\n\n\n

6. \u5bfc\u51fa\u4e0e\u90e8\u7f72<\/strong>\uff1a\u4fdd\u5b58\u6a21\u578b\uff0c\u91cf\u5316\u4f18\u5316\uff0c\u90e8\u7f72\u670d\u52a1\u3002<\/p>\n\n\n\n

2.2 \u6570\u636e\u51c6\u5907<\/strong><\/h3>\n\n\n\n

\u25cf \u6570\u636e\u683c\u5f0f<\/strong>\uff1a\u901a\u5e38\u91c7\u7528\u201c\u6307\u4ee4-\u56de\u7b54\u201d\u5bf9\uff08instruction-input-output\uff09\u6216\u5bf9\u8bdd\u683c\u5f0f\u3002<\/p>\n\n\n\n

\u25cf \u6570\u91cf\u8981\u6c42<\/strong>\uff1a\u51e0\u767e\u6761\u53ef\u770b\u5230\u521d\u6b65\u6548\u679c\uff0c\u6570\u5343\u6761\u4ee5\u4e0a\u6548\u679c\u66f4\u7a33\u5b9a\u3002<\/p>\n\n\n\n

\u25cf \u8d28\u91cf\u4f18\u5148<\/strong>\uff1a\u6e05\u6d17\u6570\u636e\uff0c\u4fdd\u8bc1\u7b54\u6848\u51c6\u786e\u3001\u98ce\u683c\u4e00\u81f4\u3002<\/p>\n\n\n\n

\u793a\u4f8b\u6570\u636e\uff08JSONL\u683c\u5f0f\uff09\uff1a<\/p>\n\n\n\n

{\"instruction\": \"\u89e3\u91ca\u4ec0\u4e48\u662fRESTful API\", \"output\": \"RESTful API\u662f\u4e00\u79cd\u57fa\u4e8eHTTP\u7684\u67b6\u6784\u98ce\u683c\uff0c\u4f7f\u7528\u8d44\u6e90\u6982\u5ff5\u548c\u6807\u51c6HTTP\u65b9\u6cd5\uff08GET\u3001POST\u3001PUT\u3001DELETE\uff09\u8fdb\u884c\u901a\u4fe1\u3002\"}
{\"instruction\": \"\u5c06\u4ee5\u4e0b\u82f1\u6587\u7ffb\u8bd1\u6210\u4e2d\u6587\uff1aHello world\", \"output\": \"\u4f60\u597d\u4e16\u754c\"}<\/code><\/pre>\n\n\n\n
2.3 \u5fae\u8c03\u65b9\u6cd5<\/strong><\/h3>\n\n\n\n
\u65b9\u6cd5<\/strong><\/td>\u8bf4\u660e<\/strong><\/td>\u4f18\u70b9<\/strong><\/td>\u7f3a\u70b9<\/strong><\/td><\/tr>
\u5168\u91cf\u5fae\u8c03<\/strong><\/td>\u66f4\u65b0\u6240\u6709\u53c2\u6570<\/td>\u6548\u679c\u6700\u597d<\/td>\u9700\u8981\u5927\u91cf\u663e\u5b58\uff0c\u8bad\u7ec3\u6162<\/td><\/tr>
LoRA<\/strong><\/td>\u4f4e\u79e9\u9002\u914d\uff0c\u53ea\u8bad\u7ec3\u5c11\u91cf\u989d\u5916\u53c2\u6570<\/td>\u663e\u5b58\u5360\u7528\u5c0f\uff0c\u8bad\u7ec3\u5feb\uff0c\u6613\u5207\u6362<\/td>\u6548\u679c\u7565\u900a\u4e8e\u5168\u91cf<\/td><\/tr>
QLoRA<\/strong><\/td>LoRA + \u91cf\u5316<\/td>4-bit\u91cf\u5316\uff0c\u5355\u5361\u53ef\u5fae\u8c03\u51e0\u5341B\u6a21\u578b<\/td>\u7cbe\u5ea6\u7565\u6709\u635f\u5931<\/td><\/tr>
Adapter<\/strong><\/td>\u63d2\u5165\u5c0f\u578b\u7f51\u7edc\u5c42<\/td>\u53c2\u6570\u5c11\uff0c\u6613\u63d2\u62d4<\/td>\u63a8\u7406\u7565\u589e\u5f00\u9500<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\u76ee\u524dLoRA\/QLoRA\u662f\u4e3b\u6d41\uff0c\u5c24\u5176\u9002\u5408\u8d44\u6e90\u6709\u9650\u7684\u73af\u5883\u3002<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u4e09\u3001\u5fae\u8c03\u5b9e\u6218\uff1a\u7528LLaMA-Factory\u5fae\u8c03Qwen<\/strong><\/h2>\n\n\n\n
LLaMA-Factory\u662f\u4e00\u4e2a\u6613\u7528\u7684\u5fae\u8c03\u6846\u67b6\uff0c\u652f\u6301\u591a\u79cd\u6a21\u578b\u548c\u65b9\u6cd5\u3002\u4ee5\u4e0b\u793a\u4f8b\u4f7f\u7528Qwen-7B\u8fdb\u884cLoRA\u5fae\u8c03\u3002<\/p>\n\n\n\n
3.1 \u73af\u5883\u51c6\u5907<\/strong><\/h3>\n\n\n\n
git clone \n            https://github.com\/hiyouga\/LLaMA-Factory.git\n<\/a>\ncd LLaMA-Factory\npip install -r \n            requirements.txt\n<\/a><\/code><\/pre>\n\n\n\n
3.2 \u6570\u636e\u51c6\u5907<\/strong><\/h3>\n\n\n\n
\u5c06\u8bad\u7ec3\u6570\u636e\u653e\u5728 data \u76ee\u5f55\uff0c\u683c\u5f0f\u4e3aJSON\uff0c\u5e76\u5728 data\/ dataset_info.json <\/a> \u4e2d\u6ce8\u518c\u3002<\/p>\n\n\n\n
3.3 \u542f\u52a8\u5fae\u8c03<\/strong><\/h3>\n\n\n\n
python src\/\n            train_bash.py\n          <\/a> \\
    --stage sft \\
    --model_name_or_path Qwen\/Qwen-7B \\
    --dataset my_dataset \\
    --template qwen \\
    --finetuning_type lora \\
    --lora_target q_proj,v_proj \\
    --output_dir .\/output \\
    --per_device_train_batch_size 4 \\
    --gradient_accumulation_steps 4 \\
    --learning_rate 5e-5 \\
    --num_train_epochs 3 \\
    --fp16<\/code><\/pre>\n\n\n\n
3.4 \u5408\u5e76\u6743\u91cd\u5e76\u5bfc\u51fa<\/strong><\/h3>\n\n\n\n
\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u5c06LoRA\u6743\u91cd\u5408\u5e76\u5230\u57fa\u5ea7\u6a21\u578b\uff0c\u4fbf\u4e8e\u90e8\u7f72\uff1a<\/p>\n\n\n\n
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = \n            AutoModelForCausalLM.from_pretrained(\"Qwen\/Qwen-7B\")\n          <\/a>
tokenizer = \n            AutoTokenizer.from_pretrained(\"Qwen\/Qwen-7B\")\n          <\/a>
model = \n            PeftModel.from_pretrained(base_model,\n          <\/a> \".\/output\")
merged_model = \n            model.merge_and_unload()\n          <\/a>
\n            merged_model.save_pretrained(\".\/merged_model\")\n          <\/a>
\n            tokenizer.save_pretrained(\".\/merged_model\")\n          <\/a><\/code><\/pre>\n\n\n\n
\u56db\u3001\u6a21\u578b\u90e8\u7f72\uff1a\u4ece\u8bad\u7ec3\u5230\u670d\u52a1<\/strong><\/h2>\n\n\n\n
\u90e8\u7f72\u7684\u76ee\u6807\u662f\u5c06\u6a21\u578b\u4ee5API\u6216\u670d\u52a1\u7684\u5f62\u5f0f\u63d0\u4f9b\u7ed9\u5176\u4ed6\u7cfb\u7edf\u8c03\u7528\u3002<\/p>\n\n\n\n
\"\u56fe\u7247\"\/<\/figure>\n\n\n\n
4.1 \u90e8\u7f72\u65b9\u5f0f\u5bf9\u6bd4<\/strong><\/h3>\n\n\n\n
\u65b9\u5f0f<\/strong><\/td>\u9002\u7528\u573a\u666f<\/strong><\/td>\u4f18\u70b9<\/strong><\/td>\u7f3a\u70b9<\/strong><\/td><\/tr>
\u6a21\u578bAPI\u670d\u52a1<\/strong><\/td>\u901a\u7528\uff0c\u5feb\u901f\u4e0a\u7ebf<\/td>\u7b80\u5355\uff0c\u65e0\u9700\u7ba1\u7406\u786c\u4ef6<\/td>\u4f9d\u8d56\u7b2c\u4e09\u65b9\uff0c\u6210\u672c\u53ef\u63a7\u6027\u5dee<\/td><\/tr>
\u81ea\u6258\u7ba1\uff08GPU\uff09<\/strong><\/td>\u9ad8\u9891\u8c03\u7528\u3001\u6570\u636e\u654f\u611f<\/td>\u5b8c\u5168\u63a7\u5236\uff0c\u957f\u671f\u6210\u672c\u4f4e<\/td>\u9700\u8fd0\u7ef4GPU\u96c6\u7fa4<\/td><\/tr>
\u8fb9\u7f18\u7aef<\/strong><\/td>\u79bb\u7ebf\u3001\u4f4e\u5ef6\u8fdf<\/td>\u65e0\u7f51\u7edc\u4f9d\u8d56<\/td>\u6a21\u578b\u9700\u538b\u7f29\uff0c\u6027\u80fd\u53d7\u9650<\/td><\/tr>
Serverless<\/strong><\/td>\u95f4\u6b47\u8c03\u7528<\/td>\u6309\u9700\u4ed8\u8d39\uff0c\u514d\u8fd0\u7ef4<\/td>\u51b7\u542f\u52a8\u5ef6\u8fdf<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
4.2 \u5e38\u7528\u90e8\u7f72\u5de5\u5177<\/strong><\/h3>\n\n\n\n
\u5de5\u5177<\/strong><\/td>\u7279\u70b9<\/strong><\/td>\u9002\u7528\u573a\u666f<\/strong><\/td><\/tr>
vLLM<\/strong><\/td>\u9ad8\u541e\u5410\u3001PagedAttention<\/td>\u9ad8\u5e76\u53d1\u751f\u4ea7\u73af\u5883<\/td><\/tr>
TGI<\/strong><\/td>Hugging Face\u51fa\u54c1\uff0c\u529f\u80fd\u5168<\/td>\u4f01\u4e1a\u7ea7\u90e8\u7f72<\/td><\/tr>
FastAPI + Transformers<\/strong><\/td>\u7b80\u5355\u7075\u6d3b<\/td>\u539f\u578b\u3001\u5185\u90e8\u5de5\u5177<\/td><\/tr>
Ollama<\/strong><\/td>\u4e00\u952e\u8fd0\u884c\uff0c\u672c\u5730\u53cb\u597d<\/td>\u5f00\u53d1\u6d4b\u8bd5<\/td><\/tr>
TensorRT-LLM<\/strong><\/td>NVIDIA\u4f18\u5316\uff0c\u6781\u81f4\u6027\u80fd<\/td>\u5927\u89c4\u6a21\u751f\u4ea7<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
4.3 \u90e8\u7f72\u5b9e\u6218\uff1a\u4f7f\u7528vLLM\u90e8\u7f72\u5fae\u8c03\u540e\u7684\u6a21\u578b<\/strong><\/h3>\n\n\n\n
\u5b89\u88c5vLLM\uff1a<\/p>\n\n\n\n
pip install vllm<\/code><\/pre>\n\n\n\n
\u542f\u52a8\u670d\u52a1\uff1a<\/p>\n\n\n\n
python -m \n            vllm.entrypoints.openai.api_server\n          <\/a> \\
    --model .\/merged_model \\
    --port 8000<\/code><\/pre>\n\n\n\n
\u8c03\u7528API\uff1a<\/p>\n\n\n\n
import openai
\n            openai.api_base\n          <\/a> = \"http:\/\/localhost:8000\/v1\"
\n            openai.api_key\n          <\/a> = \"EMPTY\"

response = \n            openai.ChatCompletion.create(\n          <\/a>
    model=\".\/merged_model\",
    messages=[{\"role\": \"user\", \"content\": \"\u89e3\u91ca\u4ec0\u4e48\u662f\u5fae\u8c03\"}]
)
print(\n            response.choices[0].message.content)\n          <\/a><\/code><\/pre>\n\n\n\n
4.4 \u90e8\u7f72\u4f18\u5316\u8981\u70b9<\/strong><\/h3>\n\n\n\n
\u25cf \u91cf\u5316<\/strong>\uff1a\u4f7f\u7528GPTQ\u3001AWQ\u7b49\u91cf\u5316\u6280\u672f\uff0c\u51cf\u5c11\u663e\u5b58\u5360\u7528\uff0c\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n\n\n\n
\u25cf \u6279\u5904\u7406<\/strong>\uff1a\u52a8\u6001\u6279\u5904\u7406\u63d0\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n
\u25cf \u7f13\u5b58<\/strong>\uff1a\u5bf9\u5e38\u89c1\u95ee\u9898\u7f13\u5b58\u7ed3\u679c\uff0c\u964d\u4f4e\u8d1f\u8f7d\u3002<\/p>\n\n\n\n
\u25cf \u6d41\u5f0f\u8f93\u51fa<\/strong>\uff1a\u63d0\u5347\u7528\u6237\u4f53\u9a8c\uff0c\u5c24\u5176\u5bf9\u4e8e\u957f\u6587\u672c\u751f\u6210\u3002<\/p>\n\n\n\n
\u4e94\u3001\u540e\u7aef\u5f00\u53d1\u89c6\u89d2\uff1a\u96c6\u6210\u5fae\u8c03\u6a21\u578b<\/strong><\/h2>\n\n\n\n
\u4f5c\u4e3a\u540e\u7aef\u5de5\u7a0b\u5e08\uff0c\u5c06\u5fae\u8c03\u6a21\u578b\u96c6\u6210\u5230\u7cfb\u7edf\u65f6\u9700\u8003\u8651\uff1a<\/p>\n\n\n\n
\u25cf API\u8bbe\u8ba1<\/strong>\uff1a\u7edf\u4e00\u5c01\u88c5\u6a21\u578b\u8c03\u7528\u63a5\u53e3\uff0c\u4fbf\u4e8e\u524d\u7aef\u548c\u670d\u52a1\u8c03\u7528\u3002<\/p>\n\n\n\n
\u25cf \u5f02\u6b65\u5904\u7406<\/strong>\uff1a\u8017\u65f6\u4efb\u52a1\u653e\u5165\u6d88\u606f\u961f\u5217\uff0c\u907f\u514d\u963b\u585e\u3002<\/p>\n\n\n\n
\u25cf \u76d1\u63a7\u4e0e\u65e5\u5fd7<\/strong>\uff1a\u8bb0\u5f55\u8bf7\u6c42\u3001\u54cd\u5e94\u3001\u8017\u65f6\uff0c\u7528\u4e8e\u8d28\u91cf\u5206\u6790\u548c\u6210\u672c\u6838\u7b97\u3002<\/p>\n\n\n\n
\u25cf \u7070\u5ea6\u53d1\u5e03<\/strong>\uff1a\u65b0\u6a21\u578b\u5148\u5c0f\u6d41\u91cf\u9a8c\u8bc1\uff0c\u518d\u5168\u91cf\u5207\u6362\u3002<\/p>\n\n\n\n
\u25cf \u7248\u672c\u7ba1\u7406<\/strong>\uff1a\u4fdd\u5b58\u591a\u4e2a\u6a21\u578b\u7248\u672c\uff0c\u652f\u6301\u56de\u6eda\u548cA\/B\u6d4b\u8bd5\u3002<\/p>\n\n\n\n
<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u4e09\u3001\u5fae\u8c03\u5b9e\u6218\uff1a\u7528LLaMA-Factory\u5fae\u8c03Qwen<\/strong><\/h2>\n\n\n\n
LLaMA-Factory\u662f\u4e00\u4e2a\u6613\u7528\u7684\u5fae\u8c03\u6846\u67b6\uff0c\u652f\u6301\u591a\u79cd\u6a21\u578b\u548c\u65b9\u6cd5\u3002\u4ee5\u4e0b\u793a\u4f8b\u4f7f\u7528Qwen-7B\u8fdb\u884cLoRA\u5fae\u8c03\u3002<\/p>\n\n\n\n
3.1 \u73af\u5883\u51c6\u5907<\/strong><\/h3>\n\n\n\n
git clone \n            https://github.com\/hiyouga\/LLaMA-Factory.git\n          <\/a>
cd LLaMA-Factory
pip install -r \n            requirements.txt\n          <\/a><\/code><\/pre>\n\n\n\n
3.2 \u6570\u636e\u51c6\u5907<\/strong><\/h3>\n\n\n\n
\u5c06\u8bad\u7ec3\u6570\u636e\u653e\u5728 data \u76ee\u5f55\uff0c\u683c\u5f0f\u4e3aJSON\uff0c\u5e76\u5728 data\/ dataset_info.json <\/a> \u4e2d\u6ce8\u518c\u3002<\/p>\n\n\n\n
3.3 \u542f\u52a8\u5fae\u8c03<\/strong><\/h3>\n\n\n\n
python src\/\n            train_bash.py\n          <\/a> \\
    --stage sft \\
    --model_name_or_path Qwen\/Qwen-7B \\
    --dataset my_dataset \\
    --template qwen \\
    --finetuning_type lora \\
    --lora_target q_proj,v_proj \\
    --output_dir .\/output \\
    --per_device_train_batch_size 4 \\
    --gradient_accumulation_steps 4 \\
    --learning_rate 5e-5 \\
    --num_train_epochs 3 \\
    --fp16<\/code><\/pre>\n\n\n\n
3.4 \u5408\u5e76\u6743\u91cd\u5e76\u5bfc\u51fa<\/strong><\/h3>\n\n\n\n
\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u5c06LoRA\u6743\u91cd\u5408\u5e76\u5230\u57fa\u5ea7\u6a21\u578b\uff0c\u4fbf\u4e8e\u90e8\u7f72\uff1a<\/p>\n\n\n\n
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = \n            AutoModelForCausalLM.from_pretrained(\"Qwen\/Qwen-7B\")\n          <\/a>
tokenizer = \n            AutoTokenizer.from_pretrained(\"Qwen\/Qwen-7B\")\n          <\/a>
model = \n            PeftModel.from_pretrained(base_model,\n          <\/a> \".\/output\")
merged_model = \n            model.merge_and_unload()\n          <\/a>
\n            merged_model.save_pretrained(\".\/merged_model\")\n          <\/a>
\n            tokenizer.save_pretrained(\".\/merged_model\")\n          <\/a><\/code><\/pre>\n\n\n\n
\u56db\u3001\u6a21\u578b\u90e8\u7f72\uff1a\u4ece\u8bad\u7ec3\u5230\u670d\u52a1<\/strong><\/h2>\n\n\n\n
\u90e8\u7f72\u7684\u76ee\u6807\u662f\u5c06\u6a21\u578b\u4ee5API\u6216\u670d\u52a1\u7684\u5f62\u5f0f\u63d0\u4f9b\u7ed9\u5176\u4ed6\u7cfb\u7edf\u8c03\u7528\u3002<\/p>\n\n\n\n
\"\u56fe\u7247\"\/<\/figure>\n\n\n\n
4.1 \u90e8\u7f72\u65b9\u5f0f\u5bf9\u6bd4<\/strong><\/h3>\n\n\n\n
\u65b9\u5f0f<\/strong><\/td>\u9002\u7528\u573a\u666f<\/strong><\/td>\u4f18\u70b9<\/strong><\/td>\u7f3a\u70b9<\/strong><\/td><\/tr>
\u6a21\u578bAPI\u670d\u52a1<\/strong><\/td>\u901a\u7528\uff0c\u5feb\u901f\u4e0a\u7ebf<\/td>\u7b80\u5355\uff0c\u65e0\u9700\u7ba1\u7406\u786c\u4ef6<\/td>\u4f9d\u8d56\u7b2c\u4e09\u65b9\uff0c\u6210\u672c\u53ef\u63a7\u6027\u5dee<\/td><\/tr>
\u81ea\u6258\u7ba1\uff08GPU\uff09<\/strong><\/td>\u9ad8\u9891\u8c03\u7528\u3001\u6570\u636e\u654f\u611f<\/td>\u5b8c\u5168\u63a7\u5236\uff0c\u957f\u671f\u6210\u672c\u4f4e<\/td>\u9700\u8fd0\u7ef4GPU\u96c6\u7fa4<\/td><\/tr>
\u8fb9\u7f18\u7aef<\/strong><\/td>\u79bb\u7ebf\u3001\u4f4e\u5ef6\u8fdf<\/td>\u65e0\u7f51\u7edc\u4f9d\u8d56<\/td>\u6a21\u578b\u9700\u538b\u7f29\uff0c\u6027\u80fd\u53d7\u9650<\/td><\/tr>
Serverless<\/strong><\/td>\u95f4\u6b47\u8c03\u7528<\/td>\u6309\u9700\u4ed8\u8d39\uff0c\u514d\u8fd0\u7ef4<\/td>\u51b7\u542f\u52a8\u5ef6\u8fdf<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
4.2 \u5e38\u7528\u90e8\u7f72\u5de5\u5177<\/strong><\/h3>\n\n\n\n
\u5de5\u5177<\/strong><\/td>\u7279\u70b9<\/strong><\/td>\u9002\u7528\u573a\u666f<\/strong><\/td><\/tr>
vLLM<\/strong><\/td>\u9ad8\u541e\u5410\u3001PagedAttention<\/td>\u9ad8\u5e76\u53d1\u751f\u4ea7\u73af\u5883<\/td><\/tr>
TGI<\/strong><\/td>Hugging Face\u51fa\u54c1\uff0c\u529f\u80fd\u5168<\/td>\u4f01\u4e1a\u7ea7\u90e8\u7f72<\/td><\/tr>
FastAPI + Transformers<\/strong><\/td>\u7b80\u5355\u7075\u6d3b<\/td>\u539f\u578b\u3001\u5185\u90e8\u5de5\u5177<\/td><\/tr>
Ollama<\/strong><\/td>\u4e00\u952e\u8fd0\u884c\uff0c\u672c\u5730\u53cb\u597d<\/td>\u5f00\u53d1\u6d4b\u8bd5<\/td><\/tr>
TensorRT-LLM<\/strong><\/td>NVIDIA\u4f18\u5316\uff0c\u6781\u81f4\u6027\u80fd<\/td>\u5927\u89c4\u6a21\u751f\u4ea7<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
4.3 \u90e8\u7f72\u5b9e\u6218\uff1a\u4f7f\u7528vLLM\u90e8\u7f72\u5fae\u8c03\u540e\u7684\u6a21\u578b<\/strong><\/h3>\n\n\n\n
\u5b89\u88c5vLLM\uff1a<\/p>\n\n\n\n
pip install vllm<\/code><\/pre>\n\n\n\n
\u542f\u52a8\u670d\u52a1\uff1a<\/p>\n\n\n\n
python -m \n            vllm.entrypoints.openai.api_server\n          <\/a> \\
    --model .\/merged_model \\
    --port 8000<\/code><\/pre>\n\n\n\n
\u8c03\u7528API\uff1a<\/p>\n\n\n\n
import openai
\n            openai.api_base\n          <\/a> = \"http:\/\/localhost:8000\/v1\"
\n            openai.api_key\n          <\/a> = \"EMPTY\"

response = \n            openai.ChatCompletion.create(\n          <\/a>
    model=\".\/merged_model\",
    messages=[{\"role\": \"user\", \"content\": \"\u89e3\u91ca\u4ec0\u4e48\u662f\u5fae\u8c03\"}]
)
print(\n            response.choices[0].message.content)\n          <\/a><\/code><\/pre>\n\n\n\n
4.4 \u90e8\u7f72\u4f18\u5316\u8981\u70b9<\/strong><\/h3>\n\n\n\n
\u25cf \u91cf\u5316<\/strong>\uff1a\u4f7f\u7528GPTQ\u3001AWQ\u7b49\u91cf\u5316\u6280\u672f\uff0c\u51cf\u5c11\u663e\u5b58\u5360\u7528\uff0c\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n\n\n\n
\u25cf \u6279\u5904\u7406<\/strong>\uff1a\u52a8\u6001\u6279\u5904\u7406\u63d0\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n
\u25cf \u7f13\u5b58<\/strong>\uff1a\u5bf9\u5e38\u89c1\u95ee\u9898\u7f13\u5b58\u7ed3\u679c\uff0c\u964d\u4f4e\u8d1f\u8f7d\u3002<\/p>\n\n\n\n
\u25cf \u6d41\u5f0f\u8f93\u51fa<\/strong>\uff1a\u63d0\u5347\u7528\u6237\u4f53\u9a8c\uff0c\u5c24\u5176\u5bf9\u4e8e\u957f\u6587\u672c\u751f\u6210\u3002<\/p>\n\n\n\n
\u4e94\u3001\u540e\u7aef\u5f00\u53d1\u89c6\u89d2\uff1a\u96c6\u6210\u5fae\u8c03\u6a21\u578b<\/strong><\/h2>\n\n\n\n
\u4f5c\u4e3a\u540e\u7aef\u5de5\u7a0b\u5e08\uff0c\u5c06\u5fae\u8c03\u6a21\u578b\u96c6\u6210\u5230\u7cfb\u7edf\u65f6\u9700\u8003\u8651\uff1a<\/p>\n\n\n\n
\u25cf API\u8bbe\u8ba1<\/strong>\uff1a\u7edf\u4e00\u5c01\u88c5\u6a21\u578b\u8c03\u7528\u63a5\u53e3\uff0c\u4fbf\u4e8e\u524d\u7aef\u548c\u670d\u52a1\u8c03\u7528\u3002<\/p>\n\n\n\n
\u25cf \u5f02\u6b65\u5904\u7406<\/strong>\uff1a\u8017\u65f6\u4efb\u52a1\u653e\u5165\u6d88\u606f\u961f\u5217\uff0c\u907f\u514d\u963b\u585e\u3002<\/p>\n\n\n\n
\u25cf \u76d1\u63a7\u4e0e\u65e5\u5fd7<\/strong>\uff1a\u8bb0\u5f55\u8bf7\u6c42\u3001\u54cd\u5e94\u3001\u8017\u65f6\uff0c\u7528\u4e8e\u8d28\u91cf\u5206\u6790\u548c\u6210\u672c\u6838\u7b97\u3002<\/p>\n\n\n\n
\u25cf \u7070\u5ea6\u53d1\u5e03<\/strong>\uff1a\u65b0\u6a21\u578b\u5148\u5c0f\u6d41\u91cf\u9a8c\u8bc1\uff0c\u518d\u5168\u91cf\u5207\u6362\u3002<\/p>\n\n\n\n
\u25cf \u7248\u672c\u7ba1\u7406<\/strong>\uff1a\u4fdd\u5b58\u591a\u4e2a\u6a21\u578b\u7248\u672c\uff0c\u652f\u6301\u56de\u6eda\u548cA\/B\u6d4b\u8bd5\u3002<\/p>\n\n\n\n
\u516d\u3001\u603b\u7ed3\u4e0e\u5c55\u671b<\/strong><\/h2>\n\n\n\n
\u6838\u5fc3\u8981\u70b9<\/strong><\/h3>\n\n\n\n
\u7ef4\u5ea6<\/strong><\/td>\u5173\u952e\u5185\u5bb9<\/strong><\/td><\/tr>
\u5fae\u8c03<\/strong><\/td>\u7528\u4e1a\u52a1\u6570\u636e\u8ba9\u6a21\u578b\u66f4\u8d34\u5408\u573a\u666f\uff0c\u5e38\u7528LoRA\/QLoRA<\/td><\/tr>
\u6570\u636e\u51c6\u5907<\/strong><\/td>\u6307\u4ee4-\u56de\u7b54\u5bf9\uff0c\u8d28\u91cf\u91cd\u4e8e\u6570\u91cf<\/td><\/tr>
\u90e8\u7f72\u5de5\u5177<\/strong><\/td>vLLM\uff08\u9ad8\u541e\u5410\uff09\u3001TGI\uff08\u4f01\u4e1a\u7ea7\uff09\u3001Ollama\uff08\u672c\u5730\uff09<\/td><\/tr>
\u4f18\u5316<\/strong><\/td>\u91cf\u5316\u3001\u6279\u5904\u7406\u3001\u7f13\u5b58\u3001\u6d41\u5f0f\u8f93\u51fa<\/td><\/tr>
\u540e\u7aef\u96c6\u6210<\/strong><\/td>API\u8bbe\u8ba1\u3001\u5f02\u6b65\u3001\u76d1\u63a7\u3001\u7070\u5ea6\u3001\u7248\u672c\u7ba1\u7406<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
<\/p>","protected":false},"excerpt":{"rendered":"
\u5df2\u7ecf\u4e86\u89e3\u4e86\u5927\u6a21\u578b\u7684\u57fa\u7840\u6982\u5ff5\u3001RAG\u3001Agent\u3001Transformer\u67b6\u6784\u7b49\u6838\u5fc3\u77e5\u8bc6\u3002\u4eca\u5929\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u5c06\u9884 […]<\/p>","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"wp-custom-template","format":"standard","meta":{"_acf_changed":false,"_crdt_document":"","_uag_custom_page_level_css":"","footnotes":""},"categories":[9],"tags":[26,23],"class_list":["post-1250","post","type-post","status-publish","format-standard","hentry","category-blog","tag-26","tag-23"],"acf":[],"views":43,"uagb_featured_image_src":{"full":false,"thumbnail":false,"medium":false,"medium_large":false,"large":false,"1536x1536":false,"2048x2048":false,"trp-custom-language-flag":false},"uagb_author_info":{"display_name":"stark, tony","author_link":"https:\/\/51ai.website\/en\/author\/admin\/"},"uagb_comment_info":6,"uagb_excerpt":"\u5df2\u7ecf\u4e86\u89e3\u4e86\u5927\u6a21\u578b\u7684\u57fa\u7840\u6982\u5ff5\u3001RAG\u3001Agent\u3001Transformer\u67b6\u6784\u7b49\u6838\u5fc3\u77e5\u8bc6\u3002\u4eca\u5929\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u5c06\u9884…","_links":{"self":[{"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/posts\/1250","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/comments?post=1250"}],"version-history":[{"count":6,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/posts\/1250\/revisions"}],"predecessor-version":[{"id":1258,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/posts\/1250\/revisions\/1258"}],"wp:attachment":[{"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/media?parent=1250"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/categories?post=1250"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/51ai.website\/en\/wp-json\/wp\/v2\/tags?post=1250"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}