app.py<\/code> (Python Flask \u5e94\u7528)<\/h3>\n\u4e00\u4e2a\u7b80\u5355\u7684 Python Flask \u5e94\u7528\u3002<\/p>\n
from flask import Flask, jsonify\nimport platform\n\napp = Flask(__name__)\n\n@app.route('\/')\ndef hello():\n return jsonify({\n "message": "\u606d\u559c\uff01\u4f60\u7684\u7b2c114514\u4e2aDocker\u5bb9\u5668\u6b63\u5728\u8fd0\u884c\uff01",\n "platform": platform.platform(),\n "python_version": platform.python_version()\n })\n\n@app.route('\/health')\ndef health():\n return jsonify({"status": "healthy"})\n\nif __name__ == '__main__':\n app.run(host='0.0.0.0', port=5000)<\/code><\/pre>\n2. requirements.txt<\/code> (Python \u4f9d\u8d56\u6587\u4ef6)<\/h3>\n\u5305\u542b Flask<\/code> \u4f9d\u8d56\u3002<\/p>\nFlask<\/code><\/pre>\n3. Dockerfile<\/code> (\u5bb9\u5668\u6784\u5efa\u6307\u4ee4)<\/h3>\n\u5b9a\u4e49\u4e86\u5982\u4f55\u6784\u5efa Python \u5e94\u7528\u955c\u50cf\u7684\u6307\u4ee4\u96c6\u3002<\/p>\n
# \u57fa\u7840\u955c\u50cf - \u5c31\u50cf\u9009\u62e9\u64cd\u4f5c\u7cfb\u7edf\nFROM python:3.10-slim\n\n# \u8bbe\u7f6e\u5de5\u4f5c\u76ee\u5f55 - \u5c31\u50cf\u8fdb\u5165\u7279\u5b9a\u6587\u4ef6\u5939\nWORKDIR \/app\n\n# \u590d\u5236\u4f9d\u8d56\u6587\u4ef6\nCOPY requirements.txt .\n\n# \u5b89\u88c5\u4f9d\u8d56 - \u5c31\u50cf\u5b89\u88c5\u8f6f\u4ef6\nRUN pip install --no-cache-dir -r requirements.txt\n\n# \u590d\u5236\u5e94\u7528\u4ee3\u7801\nCOPY app.py .\n\n# \u66b4\u9732\u7aef\u53e3 - \u544a\u8bc9Docker\u54ea\u4e2a\u7aef\u53e3\u4f1a\u88ab\u4f7f\u7528\nEXPOSE 5000\n\n# \u542f\u52a8\u547d\u4ee4 - \u5bb9\u5668\u542f\u52a8\u65f6\u6267\u884c\nCMD ["python", "app.py"]<\/code><\/pre>\n\n\n\u6838\u5fc3\u547d\u4ee4<\/strong>:<\/p>\nBash<\/p>\n
# 1. \u6784\u5efa\u955c\u50cf (\u6211\u4eec\u5c06\u955c\u50cf\u547d\u540d\u4e3a first)\ndocker build -t first .\n\n# 2. \u8fd0\u884c\u5bb9\u5668 (\u5c06\u4e3b\u673a\u76845000\u7aef\u53e3\u6620\u5c04\u5230\u5bb9\u5668\u76845000\u7aef\u53e3)\ndocker run --rm -p 5000:5000 first<\/code><\/pre>\n<\/li>\n\n\u6210\u679c<\/strong>: \u6210\u529f\u901a\u8fc7\u6d4f\u89c8\u5668\u8bbf\u95ee http:\/\/localhost:5000<\/code>\uff0c\u5e76\u770b\u5230\u4e86\u5e94\u7528\u8fd4\u56de\u7684 JSON \u795d\u8d3a\u4fe1\u606f\u3002<\/p>\n<\/li>\n<\/ul>\nLevel 2: \u642d\u5efa GPU \u52a0\u901f\u7684 Docker \u73af\u5883<\/strong><\/h2>\n\u76ee\u6807<\/strong><\/h3>\n\u6b64\u9636\u6bb5\u7684\u6838\u5fc3\u76ee\u6807\u662f\u8ba9 Docker \u5bb9\u5668\u80fd\u591f\u8c03\u7528\u4f60\u4e3b\u673a\uff08Arch Linux\uff09\u7684 NVIDIA GPU \u5f3a\u5927\u7b97\u529b<\/strong>\u3002\u8fd9\u662f\u6240\u6709\u540e\u7eed AI \u5e94\u7528\u5bb9\u5668\u5316\u90e8\u7f72\u7684\u7edd\u5bf9\u57fa\u7840\u3002<\/p>\n\u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u7ed9\u4f60\u7684 Docker\u201c\u96c6\u88c5\u7bb1\u201d\u5b89\u88c5\u4e00\u4e2a\u7279\u6b8a\u7684\u63a5\u53e3\uff0c\u8ba9\u5b83\u80fd\u591f\u8fde\u63a5\u5230\u7535\u8111\u7684\u201c\u6da1\u8f6e\u589e\u538b\u5f15\u64ce\u201d\uff08GPU\uff09\u3002<\/p>\n
\u7b2c\u4e00\u6b65\uff1a\u5b89\u88c5\u6838\u5fc3\u7ec4\u4ef6 - NVIDIA Container Toolkit<\/strong><\/h3>\n\u8981\u5b9e\u73b0 Docker \u548c GPU \u4e4b\u95f4\u7684\u901a\u4fe1\uff0c\u6211\u4eec\u9700\u8981 NVIDIA \u5b98\u65b9\u63d0\u4f9b\u7684\u4e00\u4e2a\u5173\u952e\u5de5\u5177\uff1aNVIDIA Container Toolkit\u3002<\/p>\n
\n\n\u5728 Arch Linux \u4e0a\u5b89\u88c5 Toolkit<\/strong>\n\u8fd9\u4e2a\u5de5\u5177\u5305\u5df2\u7ecf\u88ab\u6536\u5f55\u5728 Arch \u7684\u5b98\u65b9\u4ed3\u5e93\u4e2d\uff0c\u5b89\u88c5\u975e\u5e38\u7b80\u5355\u3002<\/p>\nsudo pacman -S nvidia-container-toolkit<\/code><\/pre>\n<\/li>\n\n\u91cd\u542f Docker \u670d\u52a1<\/strong>\n\u5b89\u88c5\u5b8c\u6210\u540e\uff0c\u5fc5\u987b\u91cd\u542f Docker \u670d\u52a1\uff0c\u8fd9\u6837 Docker \u624d\u80fd\u8bc6\u522b\u5230\u65b0\u5b89\u88c5\u7684 Toolkit \u5e76\u52a0\u8f7d\u5b83\u7684\u914d\u7f6e\u3002<\/p>\nsudo systemctl restart docker<\/code><\/pre>\n<\/li>\n<\/ol>\n\u7b2c\u4e8c\u6b65\uff1a\u57fa\u7840\u9a8c\u8bc1 - \u5728\u5bb9\u5668\u5185\u8fd0\u884c nvidia-smi<\/code><\/strong><\/h3>\n\u8fd9\u662f\u6700\u76f4\u63a5\u3001\u6700\u5feb\u901f\u7684\u9a8c\u8bc1\u65b9\u6cd5\uff0c\u7528\u6765\u786e\u8ba4 Toolkit \u662f\u5426\u5b89\u88c5\u6210\u529f\u3002\u6211\u4eec\u5c06\u542f\u52a8\u4e00\u4e2a\u5305\u542b CUDA \u73af\u5883\u7684\u5b98\u65b9\u5bb9\u5668\uff0c\u5e76\u5728\u5bb9\u5668\u5185\u90e8<\/strong>\u5c1d\u8bd5\u8fd0\u884c nvidia-smi<\/code> \u547d\u4ee4\u3002<\/p>\n\n\n\u8fd0\u884c\u6d4b\u8bd5\u5bb9\u5668<\/strong><\/p>\ndocker run --rm --gpus all nvidia\/cuda:12.4.1-base-ubuntu22.04 nvidia-smi<\/code><\/pre>\n\u547d\u4ee4\u89e3\u91ca<\/strong>:<\/p>\n\n--rm<\/code>: \u5bb9\u5668\u8fd0\u884c\u7ed3\u675f\u540e\u81ea\u52a8\u5220\u9664\uff0c\u4fdd\u6301\u7cfb\u7edf\u5e72\u51c0\u3002<\/li>\n--gpus all<\/code>: \u8fd9\u662f\u6700\u5173\u952e\u7684\u53c2\u6570\uff0c\u5b83\u544a\u8bc9 Docker \u628a\u4e3b\u673a\u4e0a\u6240\u6709<\/strong>\u53ef\u7528\u7684 NVIDIA GPU \u90fd\u6388\u6743\u7ed9\u8fd9\u4e2a\u5bb9\u5668\u4f7f\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n\n\u9884\u671f\u6210\u679c<\/strong>\n\u5982\u679c\u4e00\u5207\u914d\u7f6e\u6b63\u786e\uff0c\u4f60\u4f1a\u5728\u7ec8\u7aef\u770b\u5230\u548c\u4f60\u5728\u81ea\u5df1\u7535\u8111\u4e0a\u76f4\u63a5\u8fd0\u884c nvidia-smi<\/code> \u65f6\u4e00\u6a21\u4e00\u6837<\/strong>\u7684\u663e\u5361\u4fe1\u606f\u8868\u683c\u3002\u8fd9\u8bc1\u660e Docker \u5df2\u7ecf\u6210\u529f\u5730\u201c\u770b\u201d\u5230\u4e86\u4f60\u7684 GPU\u3002<\/p>\n<\/li>\n<\/ol>\n\u7b2c\u4e09\u6b65\uff1a\u8fdb\u9636\u9a8c\u8bc1 - \u5728\u5bb9\u5668\u5185\u7528 PyTorch \u68c0\u6d4b GPU<\/strong><\/h3>\nnvidia-smi<\/code> \u53ea\u80fd\u8bc1\u660e\u5e95\u5c42\u9a71\u52a8\u662f\u901a\u7684\u3002\u4e3a\u4e86\u786e\u4fdd AI \u6846\u67b6\u4e5f\u80fd\u6b63\u5e38\u4f7f\u7528 GPU\uff0c\u6211\u4eec\u53ef\u4ee5\u521b\u5efa\u4e00\u4e2a\u5305\u542b PyTorch \u7684\u73af\u5883\u6765\u505a\u4e00\u6b21\u6700\u7ec8\u6d4b\u8bd5\u3002<\/p>\n\n\n\u521b\u5efa Python \u6d4b\u8bd5\u811a\u672c (gpu_test.py<\/code>)<\/strong>\n\u5728\u4f60\u7684\u9879\u76ee\u6839\u76ee\u5f55 ~\/hunyuan-project<\/code> \u4e0b\uff0c\u521b\u5efa\u4e00\u4e2a\u540d\u4e3a gpu_test.py<\/code> \u7684\u6587\u4ef6\u3002<\/p>\ncd ~\/hunyuan-project\nnano gpu_test.py<\/code><\/pre>\n\u5c06\u4ee5\u4e0b\u4ee3\u7801\u7c98\u8d34\u8fdb\u53bb\uff1a<\/p>\n
import torch\n\nprint("--- PyTorch CUDA \u68c0\u6d4b ---")\nif torch.cuda.is_available():\n print("\u2705 \u592a\u68d2\u4e86\uff01PyTorch \u6210\u529f\u68c0\u6d4b\u5230 CUDA\uff01")\n print(f"CUDA \u8bbe\u5907\u6570\u91cf: {torch.cuda.device_count()}")\n gpu_index = 0\n print(f"\u5f53\u524d GPU \u7d22\u5f15: {gpu_index}")\n print(f"\u5f53\u524d GPU \u540d\u79f0: {torch.cuda.get_device_name(gpu_index)}")\n total_mem = torch.cuda.get_device_properties(gpu_index).total_memory \/ 1e9\n print(f"\u663e\u5b58\u603b\u91cf: {total_mem:.2f} GB")\nelse:\n print("\u274c \u7cdf\u7cd5... PyTorch \u672a\u80fd\u68c0\u6d4b\u5230 CUDA \u8bbe\u5907\u3002")\nprint("-------------------------")<\/code><\/pre>\n<\/li>\n\n\u4e3a PyTorch \u6d4b\u8bd5\u521b\u5efa\u4e13\u7528 Dockerfile<\/strong>\n\u4e3a\u4e86\u907f\u514d\u6c61\u67d3\u6211\u4eec\u4e4b\u524d\u7684 vllm-service<\/code> \u955c\u50cf\uff0c\u6211\u4eec\u521b\u5efa\u4e00\u4e2a\u65b0\u7684 Dockerfile\uff0c\u4e13\u95e8\u7528\u4e8e\u8fd9\u4e2a\u6d4b\u8bd5\u3002\u6211\u4eec\u628a\u5b83\u547d\u540d\u4e3a Dockerfile.gpu-test<\/code>\u3002<\/p>\nnano Dockerfile.gpu-test<\/code><\/pre>\n\u5c06\u4ee5\u4e0b\u5185\u5bb9\u7c98\u8d34\u8fdb\u53bb\uff1a<\/p>\n
# \u76f4\u63a5\u4f7f\u7528 PyTorch \u5b98\u65b9\u63d0\u4f9b\u7684\u3001\u5185\u7f6e CUDA \u548c CUDNN \u7684\u955c\u50cf\uff0c\u8fd9\u662f\u6700\u65b9\u4fbf\u53ef\u9760\u7684\u65b9\u5f0f\nFROM pytorch\/pytorch:2.3.1-cuda12.1-cudnn8-runtime\n\n# \u8bbe\u7f6e\u5de5\u4f5c\u76ee\u5f55\nWORKDIR \/app\n\n# \u5c06\u6211\u4eec\u7684\u6d4b\u8bd5\u811a\u672c\u590d\u5236\u8fdb\u53bb\nCOPY gpu_test.py .\n\n# \u8bbe\u7f6e\u5bb9\u5668\u542f\u52a8\u65f6\u8981\u6267\u884c\u7684\u547d\u4ee4\nCMD ["python", "gpu_test.py"]<\/code><\/pre>\n<\/li>\n\n\u6784\u5efa\u5e76\u8fd0\u884c PyTorch \u6d4b\u8bd5\u955c\u50cf<\/strong><\/p>\n# \u4f7f\u7528 -f \u53c2\u6570\u6307\u5b9a\u6211\u4eec\u521a\u521a\u521b\u5efa\u7684 Dockerfile \u6587\u4ef6\u540d\u6765\u6784\u5efa\u955c\u50cf\ndocker build -t gpu-test-app -f Dockerfile.gpu-test .\n\n# \u8fd0\u884c\u8fd9\u4e2a\u6d4b\u8bd5\u5bb9\u5668\uff0c\u540c\u6837\u9700\u8981 --gpus all \u53c2\u6570\ndocker run --rm --gpus all gpu-test-app<\/code><\/pre>\n<\/li>\n\n\u9884\u671f\u6210\u679c<\/strong>\n\u8fd0\u884c\u6210\u529f\u540e\uff0c\u4f60\u4f1a\u5728\u7ec8\u7aef\u770b\u5230\u7c7b\u4f3c\u4e0b\u9762\u7684\u8f93\u51fa\uff1a<\/p>\n--- PyTorch CUDA \u68c0\u6d4b ---\n\u2705 \u592a\u68d2\u4e86\uff01PyTorch \u6210\u529f\u68c0\u6d4b\u5230 CUDA\uff01\nCUDA \u8bbe\u5907\u6570\u91cf: 1\n\u5f53\u524d GPU \u7d22\u5f15: 0\n\u5f53\u524d GPU \u540d\u79f0: NVIDIA GeForce RTX 4060 Laptop GPU\n\u663e\u5b58\u603b\u91cf: 7.99 GB\n-------------------------<\/code><\/pre>\n<\/li>\n<\/ol>\n\u4e09\u3001 Level 3: \u90e8\u7f72 vLLM \u63a8\u7406\u670d\u52a1<\/h2>\n \u6b64\u9636\u6bb5\u662f\u9879\u76ee\u7684\u6838\u5fc3\uff0c\u76ee\u6807\u662f\u6784\u5efa\u5e76\u8fd0\u884c\u4e00\u4e2a\u529f\u80fd\u5f3a\u5927\u7684 vLLM \u670d\u52a1\u5bb9\u5668\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u9047\u5230\u4e86\u591a\u6b21\u4f9d\u8d56\u548c\u73af\u5883\u4e0d\u517c\u5bb9\u95ee\u9898\uff0c\u6700\u7ec8\u901a\u8fc7\u4e00\u4e2a\u7a33\u5065\u7684 Dockerfile \u65b9\u6848\u5f97\u4ee5\u89e3\u51b3\u3002<\/p>\n
Dockerfile<\/p>\n
# \u4f7f\u7528\u4e00\u4e2a\u975e\u5e38\u65b0\u7684 CUDA 12.4.1 \u57fa\u7840\u955c\u50cf\uff0c\u4ee5\u83b7\u5f97\u6700\u597d\u7684\u517c\u5bb9\u6027\nFROM nvidia\/cuda:12.4.1-base-ubuntu22.04\n\n# \u8bbe\u7f6e\u73af\u5883\u53d8\u91cf\uff0c\u907f\u514d apt-get \u5728\u6784\u5efa\u65f6\u5f39\u51fa\u4ea4\u4e92\u7a97\u53e3\nENV DEBIAN_FRONTEND=noninteractive\n\n# \u76f4\u63a5\u4e0b\u8f7d\u5e76\u4ece\u6e90\u7801\u7f16\u8bd1\u5b89\u88c5 Python 3.12\uff0c\u4ee5\u7ed5\u8fc7 PPA \u7f51\u7edc\u95ee\u9898\nRUN apt-get update && \\\n apt-get install -y wget build-essential libssl-dev zlib1g-dev libncurses5-dev \\\n libncursesw5-dev libreadline-dev libsqlite3-dev libgdbm-dev libdb5.3-dev \\\n libbz2-dev libexpat1-dev liblzma-dev tk-dev libffi-dev git && \\\n wget https:\/\/www.python.org\/ftp\/python\/3.12.4\/Python-3.12.4.tgz && \\\n tar -xf Python-3.12.4.tgz && \\\n cd Python-3.12.4 && \\\n .\/configure --enable-optimizations && \\\n make -j $(nproc) && \\\n make altinstall && \\\n cd .. && \\\n rm -rf Python-3.12.4.tgz Python-3.12.4 && \\\n apt-get clean && \\\n rm -rf \/var\/lib\/apt\/lists\/*\n\n# \u4e3a python3 \u548c pip3 \u521b\u5efa\u8f6f\u94fe\u63a5\uff0c\u6307\u5411\u6211\u4eec\u65b0\u5b89\u88c5\u7684 3.12 \u7248\u672c\nRUN ln -s \/usr\/local\/bin\/python3.12 \/usr\/local\/bin\/python3 && \\\n ln -s \/usr\/local\/bin\/pip3.12 \/usr\/local\/bin\/pip3\n\n# \u5b89\u88c5\u4e00\u7ec4\u5df2\u77e5\u517c\u5bb9\u7684\u6700\u65b0\u6838\u5fc3\u5e93\nRUN pip3 install \\\n "torch==2.4.0" \\\n "vllm==0.5.3.post1" \\\n "transformers==4.42.4" \\\n "outlines==0.0.34"\n\n# \u8bbe\u7f6e\u5de5\u4f5c\u76ee\u5f55\u548c\u6a21\u578b\u7f13\u5b58\u76ee\u5f55\nWORKDIR \/app\nRUN mkdir -p \/models\nENV TRANSFORMERS_CACHE=\/models\nEXPOSE 8000\n\n# \u542f\u52a8 vLLM \u670d\u52a1\uff0c\u4f7f\u7528\u5c0f\u7684\u6d4b\u8bd5\u6a21\u578b\nCMD ["python3", "-m", "vllm.entrypoints.openai.api_server", \\\n "--model", "facebook\/opt-125m", \\\n "--host", "0.0.0.0", \\\n "--port", "8000"]<\/code><\/pre>\n\u6838\u5fc3\u547d\u4ee4\u4e0e\u6210\u679c<\/h3>\n\n\n\u6784\u5efa\u4e0e\u8fd0\u884c<\/strong>:<\/p>\nBash<\/p>\n
# \u6784\u5efa vLLM \u670d\u52a1\u955c\u50cf\ndocker build -t vllm-service .\n\n# \u5728\u540e\u53f0\u4ee5 GPU \u6a21\u5f0f\u542f\u52a8\u5bb9\u5668\uff0c\u5e76\u6302\u8f7d\u6a21\u578b\u5b58\u50a8\u5377\ndocker run --gpus all \\\n -p 8000:8000 \\\n -v $(pwd)\/models:\/models \\\n --name vllm-server \\\n -d \\\n vllm-service<\/code><\/pre>\n<\/li>\n\n\u9a8c\u8bc1<\/strong>:<\/p>\n\n\u901a\u8fc7 docker logs vllm-server<\/code> \u786e\u8ba4\u670d\u52a1\u6210\u529f\u542f\u52a8\u3002<\/li>\n\u7f16\u5199\u5e76\u8fd0\u884c test_api.py<\/code> \u811a\u672c\uff0c\u6210\u529f\u5411 http:\/\/localhost:8000<\/code> \u53d1\u9001\u8bf7\u6c42\u5e76\u6536\u5230\u4e86\u7531 facebook\/opt-125m<\/code> \u6a21\u578b\u751f\u6210\u7684\u6587\u672c\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\u9879\u76ee\u5b9e\u8df5\u8bb0\u5f55\uff1aLevel 4 - \u5927\u6a21\u578b\u90e8\u7f72\u603b\u7ed3<\/strong><\/h3>\n1. \u6838\u5fc3\u601d\u8def<\/strong><\/h4>\n\u7531\u4e8e\u672c\u849f\u84bb\u7684 RTX 4060 \u663e\u5361\u6709 8GB \u663e\u5b58\uff0c\u6240\u4ee5\u9009\u62e9\u4e86\u4e00\u4e2a\u5f88\u5c0f\u3001\u9002\u5408\u4e2a\u4eba\u7535\u8111\u90e8\u7f72\u7684 Qwen2-0.5B-Instruct<\/strong> \u6a21\u578b\u3002<\/p>\n\n\u6a21\u578b\u9009\u62e9<\/strong>\uff1aQwen2-0.5B-Instruct<\/strong> \u662f\u4e00\u4e2a\u53c2\u6570\u91cf\u4ec5\u4e3a 5 \u4ebf\u7684\u6a21\u578b\uff0c\u5b83\u5bf9\u786c\u4ef6\u8d44\u6e90\u7684\u9700\u6c42\u6781\u4f4e\uff0c\u80fd\u591f\u7a33\u5b9a\u3001\u5feb\u901f\u5730\u5728\u8bbe\u5907\u4e0a\u8fd0\u884c\u3002<\/li>\n\u90e8\u7f72\u5de5\u5177<\/strong>\uff1a\u7ee7\u7eed\u4f7f\u7528 Docker<\/strong> \u548c vLLM<\/strong>\u3002\u901a\u8fc7\u7cbe\u7b80\u7684\u542f\u52a8\u547d\u4ee4\uff0c\u6211\u4eec\u907f\u514d\u4e86\u590d\u6742\u7684\u663e\u5b58\u53c2\u6570\u8c03\u8bd5\uff0c\u8ba9\u90e8\u7f72\u8fc7\u7a0b\u53d8\u5f97\u66f4\u52a0\u7b80\u5355\u548c\u53ef\u9760\u3002<\/li>\n<\/ul>\n2. \u90e8\u7f72\u8be6\u7ec6\u6b65\u9aa4<\/strong><\/h4>\n\u8fd9\u662f\u4e00\u4e2a\u4ece\u96f6\u5f00\u59cb\u90e8\u7f72\u8fd9\u4e2a\u65b0\u6a21\u578b\u7684\u5b8c\u6574\u6307\u5357\u3002<\/p>\n
\u7b2c1\u6b65\uff1a\u4e0b\u8f7d Qwen2 0.5B \u6a21\u578b<\/strong><\/p>\n\u9996\u5148\uff0c\u505c\u6b62\u4f60\u5f53\u524d\u6b63\u5728\u8fd0\u884c\u7684 Docker \u5bb9\u5668\uff0c\u7136\u540e\u4e0b\u8f7d\u65b0\u7684\u6a21\u578b\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u6bd4\u4e4b\u524d\u5feb\u5f97\u591a\uff0c\u4e5f\u66f4\u7a33\u5b9a\u3002<\/p>\n
Bash<\/p>\n
# \u514b\u9686 Qwen2 0.5B \u6a21\u578b\ngit clone https:\/\/huggingface.co\/Qwen\/Qwen2-0.5B-Instruct<\/code><\/pre>\n\u7b2c2\u6b65\uff1a\u542f\u52a8 vLLM \u670d\u52a1<\/strong><\/p>\n\u65b0\u6a21\u578b\u7684\u542f\u52a8\u547d\u4ee4\u975e\u5e38\u7b80\u6d01\u3002\u4f60\u4e0d\u518d\u9700\u8981 --quantization<\/code>\u3001--max-model-len<\/code> \u6216 --gpu-memory-utilization<\/code> \u7b49\u590d\u6742\u53c2\u6570\u3002<\/p>\nBash<\/p>\n
docker run --gpus all --rm \\\n -p 8000:8000 \\\n -v $(pwd)\/Qwen2-0.5B-Instruct:\/models \\\n --name vllm-server \\\n vllm\/vllm-openai \\\n --model \/models \\\n --served-model-name qwen2-0.5b-instruct<\/code><\/pre>\n\u7b2c3\u6b65\uff1a\u9a8c\u8bc1\u670d\u52a1\u4e0e\u6a21\u578b<\/strong><\/p>\n\u670d\u52a1\u542f\u52a8\u540e\uff0c\u6253\u5f00\u4e00\u4e2a\u65b0\u7684\u7ec8\u7aef\uff0c\u4f7f\u7528 curl<\/code> \u547d\u4ee4\u9a8c\u8bc1 API \u63a5\u53e3\u3002<\/p>\n\n\n\u9a8c\u8bc1\u670d\u52a1\u5c31\u7eea<\/strong>\uff1a<\/p>\nBash<\/p>\n
curl http:\/\/localhost:8000\/v1\/models<\/code><\/pre>\n\u4f60\u5e94\u8be5\u4f1a\u770b\u5230\u4e00\u4e2a JSON \u54cd\u5e94\uff0c\u5176\u4e2d\u5305\u542b\u6a21\u578b ID qwen2-0.5b-instruct<\/code>\u3002<\/p>\n<\/li>\n\n\u6d4b\u8bd5\u7ffb\u8bd1\u529f\u80fd\uff1a<\/p>\n
\u4f7f\u7528\u5bf9\u8bdd\u683c\u5f0f\u5411\u6a21\u578b\u53d1\u9001\u4e00\u4e2a\u7ffb\u8bd1\u8bf7\u6c42\u3002<\/p>\n
Bash<\/p>\n
curl http:\/\/localhost:8000\/v1\/chat\/completions \\\n-H "Content-Type: application\/json" \\\n-d '{\n "model": "qwen2-0.5b-instruct",\n "messages": [\n {\n "role": "system",\n "content": "You are a professional translator. Translate all user text into English."\n },\n {\n "role": "user",\n "content": "\u4f60\u597d\uff0c\u4e16\u754c"\n }\n ],\n "max_tokens": 50\n}'<\/code><\/pre>\n\u5982\u679c\u4e00\u5207\u6b63\u5e38\uff0c\u6a21\u578b\u5c06\u8fd4\u56de\u6b63\u786e\u7684\u7ffb\u8bd1\u7ed3\u679c\uff0c\u8bc1\u660e\u4f60\u5df2\u7ecf\u6210\u529f\u90e8\u7f72\u4e86\u670d\u52a1\u3002<\/p>\n<\/li>\n<\/ul>\n
\u9879\u76ee\u8fdb\u9636\u6559\u7a0b (Level 5-7)<\/strong><\/h2>\n\u6211\u4eec\u5df2\u7ecf\u6210\u529f\u90e8\u7f72\u4e86 vLLM \u540e\u7aef\u670d\u52a1\u3002\u63a5\u4e0b\u6765\u7684\u76ee\u6807\u662f\u5c06\u8fd9\u4e2a\u5f3a\u5927\u7684 AI \u80fd\u529b\uff0c\u5c01\u88c5\u6210\u4e00\u4e2a\u7528\u6237\u53ef\u4ee5\u76f4\u89c2\u4ea4\u4e92\u7684\u5168\u529f\u80fd Web \u5e94\u7528\u3002<\/p>\n
\u51c6\u5907\u5de5\u4f5c\uff1a\u6700\u7ec8\u7684\u9879\u76ee\u7ed3\u6784<\/strong><\/h3>\n\u4e3a\u4e86\u8ba9\u9879\u76ee\u6e05\u6670\u6709\u5e8f\uff0c\u6211\u4eec\u5148\u89c4\u5212\u597d\u6700\u7ec8\u7684\u76ee\u5f55\u7ed3\u6784\u3002\u8bf7\u5728\u4f60\u7684 hunyuan-project<\/code> \u6839\u76ee\u5f55\u4e0b\uff0c\u7ec4\u7ec7\u6216\u521b\u5efa\u5982\u4e0b\u7ed3\u6784\uff1a<\/p>\nhunyuan-project\/\n\u251c\u2500\u2500 backend\/\n\u2502 \u251c\u2500\u2500 app.py # \u6211\u4eec\u7684 Flask \u540e\u7aef\uff08API \u7f51\u5173 + \u524d\u7aef\u670d\u52a1\uff09\n\u2502 \u2514\u2500\u2500 Dockerfile # \u7528\u4e8e\u6784\u5efa Flask \u540e\u7aef\u7684 Dockerfile\n\u251c\u2500\u2500 frontend\/\n\u2502 \u2514\u2500\u2500 index.html # \u7528\u6237\u770b\u5230\u7684\u7f51\u9875\u524d\u7aef\n\u251c\u2500\u2500 models\/\n\u2502 \u2514\u2500\u2500 ... # \u4f60\u4e0b\u8f7d\u7684\u6a21\u578b\u6587\u4ef6\n\u2514\u2500\u2500 docker-compose.yml # \u7f16\u6392\u6240\u6709\u670d\u52a1\u7684“\u603b\u6307\u6325”<\/code><\/pre>\n(\u8bf7\u6ce8\u610f\uff0c\u6211\u4eec\u5c06\u628a Flask \u540e\u7aef\u7684\u4ee3\u7801\u653e\u5728 backend<\/code> \u76ee\u5f55\uff0c\u524d\u7aef\u9875\u9762\u653e\u5728 frontend<\/code> \u76ee\u5f55)<\/em><\/p>\nLevel 5 & 6: \u6784\u5efa\u524d\u540e\u7aef\u5e94\u7528<\/strong><\/h3>\n\u5728\u8fd9\u4e2a\u9636\u6bb5\uff0c\u6211\u4eec\u5c06\u540c\u65f6\u5b8c\u6210 Level 5 (\u521b\u5efa\u540e\u7aef) \u548c Level 6 (\u521b\u5efa\u524d\u7aef)\uff0c\u56e0\u4e3a\u5b83\u4eec\u662f\u7d27\u5bc6\u534f\u4f5c\u7684\u3002<\/p>\n
\u7b2c\u4e00\u90e8\u5206\uff1a\u521b\u5efa\u540e\u7aef\u670d\u52a1 (backend\/app.py<\/code>)<\/strong><\/h4>\n\u6211\u4eec\u7684\u540e\u7aef\u662f\u4e00\u4e2a Python Flask \u5e94\u7528\u3002\u5b83\u6709\u4e24\u4e2a\u6838\u5fc3\u804c\u8d23\uff1a<\/p>\n
\n\u4f5c\u4e3a\u4e00\u4e2a API \u7f51\u5173<\/strong>\uff0c\u63a5\u6536\u524d\u7aef\u7684\u7ffb\u8bd1\u8bf7\u6c42\uff0c\u8f6c\u53d1\u7ed9 vLLM \u670d\u52a1\uff0c\u5e76\u628a\u7ed3\u679c\u6d41\u5f0f\u8fd4\u56de\u3002<\/li>\n\u4f5c\u4e3a\u4e00\u4e2a Web \u670d\u52a1\u5668<\/strong>\uff0c\u5411\u7528\u6237\u7684\u6d4f\u89c8\u5668\u63d0\u4f9b\u524d\u7aef\u7684 index.html<\/code> \u9875\u9762\u3002<\/li>\n<\/ol>\n1. \u521b\u5efa backend\/app.py<\/code> \u6587\u4ef6<\/strong><\/p>\nBash<\/p>\n
# \u786e\u4fdd\u5728\u9879\u76ee\u6839\u76ee\u5f55\ncd ~\/hunyuan-project\n# \u521b\u5efa backend \u76ee\u5f55\nmkdir -p backend\n# \u521b\u5efa\u5e76\u7f16\u8f91 app.py\nnano backend\/app.py<\/code><\/pre>\n2. \u7f16\u5199 app.py<\/code> \u5b8c\u6574\u4ee3\u7801<\/strong><\/p>\n\u8bf7\u5c06\u4ee5\u4e0b\u5b8c\u6574\u4ee3\u7801\u590d\u5236\u5230 app.py<\/code> \u6587\u4ef6\u4e2d\u3002\u4ee3\u7801\u7684\u8be6\u7ec6\u89e3\u91ca\u5728\u6ce8\u91ca\u91cc\u3002<\/p>\n
![\"1758640771-5f72c9db6f138cc33eb344ce877db07d\"](\"https:\/\/caoxunyi.cn\/wp-content\/uploads\/2025\/09\/1758640771-5f72c9db6f138cc33eb344ce877db07d.png\")
<\/figure>\n![\"1758640786-41e3fbaacb7e41bcef5a156e403011f9\"](\"https:\/\/caoxunyi.cn\/wp-content\/uploads\/2025\/09\/1758640786-41e3fbaacb7e41bcef5a156e403011f9.png\")
<\/figure>\n![\"1758640793-1fb975c9fde535b179f977a4b913091c\"](\"https:\/\/caoxunyi.cn\/wp-content\/uploads\/2025\/09\/1758640793-1fb975c9fde535b179f977a4b913091c-1024x513.png\")
<\/figure>\n![\"1758640792-1afa7e393fcb6cd4a5536ff51e6024fb\"](\"https:\/\/caoxunyi.cn\/wp-content\/uploads\/2025\/09\/1758640792-1afa7e393fcb6cd4a5536ff51e6024fb.png\")
<\/figure>\n