\u60f3\u8c61\u4e00\u4e0b，\u4f60\u9700\u8981\u572810\u53f0\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u4e00\u4e2a\u5b8c\u6574\u7684AI\u804a\u5929\u7cfb\u7edf\u3002\u6bcf\u53f0\u670d\u52a1\u5668\u90fd\u8981\u5b89\u88c5Python\u73af\u5883\u3001\u4e0b\u8f7d\u51e0\u5341GB\u7684\u6a21\u578b\u6587\u4ef6\u3001\u914d\u7f6evLLM\u63a8\u7406\u5f15\u64ce\u3001\u8bbe\u7f6e\u53cd\u5411\u4ee3\u7406\u3001\u90e8\u7f72\u524d\u7aef\u754c\u9762\u2026\u2026\u5982\u679c\u624b\u52a8\u64cd\u4f5c，\u4e00\u53f0\u670d\u52a1\u5668\u53ef\u80fd\u5c31\u8981\u6298\u817e\u534a\u5929，10\u53f0\u670d\u52a1\u5668\u7b80\u76f4\u662f\u4e2a\u5669\u68a6\u3002<\/p>\n
\u8fd9\u5c31\u662f\u6211\u4eec\u4eca\u5929\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002\u6211\u5c06\u5206\u4eab\u4e00\u4e2a\u57fa\u4e8eAnsible\u7684\u81ea\u52a8\u5316\u90e8\u7f72\u65b9\u6848，\u8ba9\u4f60\u7528\u4e00\u6761\u547d\u4ee4\u5c31\u80fd\u572810\u53f0\u3001\u751a\u81f3100\u53f0\u670d\u52a1\u5668\u4e0a\u6279\u91cf\u90e8\u7f72Qwen3-VL-8B AI\u804a\u5929\u7cfb\u7edf\u3002\u65e0\u8bba\u4f60\u662f\u4f01\u4e1aIT\u8fd0\u7ef4\u3001AI\u5b9e\u9a8c\u5ba4\u7ba1\u7406\u5458，\u8fd8\u662f\u9700\u8981\u7ba1\u7406\u591a\u53f0GPU\u670d\u52a1\u5668\u7684\u5f00\u53d1\u8005，\u8fd9\u5957\u65b9\u6848\u90fd\u80fd\u5e2e\u4f60\u8282\u7701\u5927\u91cf\u65f6\u95f4\u548c\u7cbe\u529b\u3002<\/p>\n
\u901a\u8fc7\u672c\u6559\u7a0b，\u4f60\u5c06\u5b66\u4f1a：<\/p>\n
\u5728\u5f00\u59cb\u81ea\u52a8\u5316\u90e8\u7f72\u4e4b\u524d，\u6211\u4eec\u5148\u4e86\u89e3\u4e00\u4e0b\u8981\u90e8\u7f72\u7684\u7cfb\u7edf\u67b6\u6784\u3002\u8fd9\u4e2aAI\u804a\u5929\u7cfb\u7edf\u5305\u542b\u4e09\u4e2a\u6838\u5fc3\u7ec4\u4ef6：<\/p>\n
\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 \u6d4f\u89c8\u5668\u5ba2\u6237\u7aef \u2502
\n\u2502 (chat.html) \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2502 HTTP
\n \u2193
\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 \u4ee3\u7406\u670d\u52a1\u5668 \u2502
\n\u2502 (proxy_server) \u2502 \u2190 \u7aef\u53e3 8000
\n\u2502 – \u9759\u6001\u6587\u4ef6\u670d\u52a1 \u2502
\n\u2502 – API \u8bf7\u6c42\u8f6c\u53d1 \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518
\n \u2502 HTTP
\n \u2193
\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 vLLM \u63a8\u7406\u5f15\u64ce \u2502 \u2190 \u7aef\u53e3 3001
\n\u2502 – \u6a21\u578b\u52a0\u8f7d \u2502
\n\u2502 – \u63a8\u7406\u8ba1\u7b97 \u2502
\n\u2502 – OpenAI API \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\u7ec4\u4ef6\u8bf4\u660e：<\/p>\n
\u5728\u591a\u53f0\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u8fd9\u4e2a\u7cfb\u7edf，\u4f1a\u9047\u5230\u51e0\u4e2a\u5178\u578b\u95ee\u9898：<\/p>\n
\u73af\u5883\u5dee\u5f02\u95ee\u9898<\/p>\n
\u6a21\u578b\u4e0b\u8f7d\u95ee\u9898<\/p>\n
\u914d\u7f6e\u7ba1\u7406\u95ee\u9898<\/p>\n
\u8fd0\u7ef4\u7ba1\u7406\u95ee\u9898<\/p>\n
Ansible\u662f\u76ee\u524d\u6700\u6d41\u884c\u7684\u81ea\u52a8\u5316\u8fd0\u7ef4\u5de5\u5177\u4e4b\u4e00，\u7279\u522b\u9002\u5408\u6279\u91cf\u90e8\u7f72\u573a\u666f：<\/p>\n
\u65e0\u4ee3\u7406\u67b6\u6784：\u4e0d\u9700\u8981\u5728\u76ee\u6807\u670d\u52a1\u5668\u4e0a\u5b89\u88c5\u5ba2\u6237\u7aef，\u901a\u8fc7SSH\u5c31\u80fd\u7ba1\u7406 \u58f0\u660e\u5f0f\u8bed\u6cd5：\u7528YAML\u7f16\u5199Playbook，\u914d\u7f6e\u5373\u4ee3\u7801 \u5e42\u7b49\u6027：\u591a\u6b21\u6267\u884c\u4e0d\u4f1a\u4ea7\u751f\u526f\u4f5c\u7528，\u786e\u4fdd\u90e8\u7f72\u4e00\u81f4\u6027 \u6a21\u5757\u4e30\u5bcc：\u5185\u7f6e\u5927\u91cf\u6a21\u5757，\u8986\u76d6\u7cfb\u7edf\u7ba1\u7406\u3001\u8f6f\u4ef6\u5b89\u88c5\u3001\u6587\u4ef6\u64cd\u4f5c\u7b49<\/p>\n
\u6211\u4eec\u7684\u81ea\u52a8\u5316\u90e8\u7f72\u6d41\u7a0b\u5206\u4e3a\u56db\u4e2a\u9636\u6bb5：<\/p>\n
\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 \u73af\u5883\u51c6\u5907\u9636\u6bb5 \u2502 \u2502 \u4f9d\u8d56\u5b89\u88c5\u9636\u6bb5 \u2502 \u2502 \u7cfb\u7edf\u90e8\u7f72\u9636\u6bb5 \u2502 \u2502 \u9a8c\u8bc1\u6d4b\u8bd5\u9636\u6bb5 \u2502
\n\u2502 1. \u7cfb\u7edf\u68c0\u67e5 \u2502\u2500\u2500\u2500\u25b6\u2502 2. \u5b89\u88c5Python \u2502\u2500\u2500\u2500\u25b6\u2502 3. \u4e0b\u8f7d\u6a21\u578b \u2502\u2500\u2500\u2500\u25b6\u2502 4. \u542f\u52a8\u670d\u52a1 \u2502
\n\u2502 – \u7528\u6237\u6743\u9650 \u2502 \u2502 – pip\u5305 \u2502 \u2502 – \u914d\u7f6e\u6587\u4ef6 \u2502 \u2502 – \u7aef\u53e3\u68c0\u67e5 \u2502
\n\u2502 – \u78c1\u76d8\u7a7a\u95f4 \u2502 \u2502 – CUDA\u5de5\u5177 \u2502 \u2502 – \u524d\u7aef\u6587\u4ef6 \u2502 \u2502 – \u670d\u52a1\u72b6\u6001 \u2502
\n\u2502 – \u7f51\u7edc\u8fde\u63a5 \u2502 \u2502 – \u7cfb\u7edf\u4f9d\u8d56 \u2502 \u2502 – \u542f\u52a8\u811a\u672c \u2502 \u2502 – \u529f\u80fd\u6d4b\u8bd5 \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518 \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518 \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518 \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u670d\u52a1\u5668\u6e05\u5355\u6587\u4ef6\u6765\u7ba1\u7406\u6240\u6709\u8981\u90e8\u7f72\u7684\u673a\u5668\u3002\u6839\u636e\u670d\u52a1\u5668\u914d\u7f6e，\u53ef\u4ee5\u5206\u4e3a\u4e0d\u540c\u7c7b\u578b：<\/p>\n
# inventory.yml
\nall:
\n vars:
\n ansible_user: root
\n ansible_ssh_private_key_file: ~\/.ssh\/id_rsa<\/p>\n
children:
\n gpu_servers: # GPU\u670d\u52a1\u5668\u7ec4
\n hosts:
\n gpu-server-01:
\n ansible_host: 192.168.1.101
\n gpu_memory: 24GB
\n gpu-server-02:
\n ansible_host: 192.168.1.102
\n gpu_memory: 16GB<\/p>\n
cpu_servers: # CPU\u670d\u52a1\u5668\u7ec4（\u7528\u4e8e\u4ee3\u7406\u670d\u52a1）
\n hosts:
\n proxy-server-01:
\n ansible_host: 192.168.1.201
\n proxy-server-02:
\n ansible_host: 192.168.1.202<\/p>\n
all_servers: # \u6240\u6709\u670d\u52a1\u5668
\n children:
\n – gpu_servers
\n – cpu_servers<\/p>\n
\u521b\u5efa\u4e00\u4e2a\u4e3b\u90e8\u7f72Playbook\u6587\u4ef6 deploy_qwen.yml：<\/p>\n
—
\n– name: \u6279\u91cf\u90e8\u7f72Qwen3-VL-8B AI\u804a\u5929\u7cfb\u7edf
\n hosts: all_servers
\n become: yes
\n vars:
\n project_dir: \/root\/build
\n model_id: "qwen\/Qwen2-VL-7B-Instruct-GPTQ-Int4"
\n model_name: "Qwen3-VL-8B-Instruct-4bit-GPTQ"
\n vllm_port: 3001
\n web_port: 8000<\/p>\n
tasks:
\n – name: \u68c0\u67e5\u7cfb\u7edf\u73af\u5883
\n block:
\n – name: \u68c0\u67e5\u64cd\u4f5c\u7cfb\u7edf\u7248\u672c
\n command: cat \/etc\/os-release
\n register: os_info<\/p>\n
– name: \u68c0\u67e5\u78c1\u76d8\u7a7a\u95f4
\n shell: df -h \/ | tail -1 | awk '{print $4}'
\n register: disk_space<\/p>\n
– name: \u68c0\u67e5GPU\u72b6\u6001（\u4ec5GPU\u670d\u52a1\u5668）
\n shell: |
\n if command -v nvidia-smi &> \/dev\/null; then
\n nvidia-smi –query-gpu=name,memory.total –format=csv,noheader
\n else
\n echo "No GPU detected"
\n fi
\n when: "'gpu_servers' in group_names"
\n register: gpu_info<\/p>\n
– name: \u663e\u793a\u68c0\u67e5\u7ed3\u679c
\n debug:
\n msg: |
\n \u64cd\u4f5c\u7cfb\u7edf: {{ os_info.stdout }}
\n \u78c1\u76d8\u7a7a\u95f4: {{ disk_space.stdout }}
\n GPU\u4fe1\u606f: {{ gpu_info.stdout if gpu_info is defined else 'N\/A' }}<\/p>\n
– name: \u5b89\u88c5\u7cfb\u7edf\u4f9d\u8d56
\n apt:
\n name:
\n – python3-pip
\n – python3-venv
\n – git
\n – wget
\n – curl
\n – supervisor
\n state: present
\n update_cache: yes
\n when: ansible_os_family == "Debian"<\/p>\n
– name: \u4e3aCentOS\u5b89\u88c5\u4f9d\u8d56
\n yum:
\n name:
\n – python3-pip
\n – git
\n – wget
\n – curl
\n – supervisor
\n state: present
\n when: ansible_os_family == "RedHat"<\/p>\n
– name: \u521b\u5efa\u9879\u76ee\u76ee\u5f55
\n file:
\n path: "{{ project_dir }}"
\n state: directory
\n mode: '0755'<\/p>\n
\u6a21\u578b\u4e0b\u8f7d\u662f\u90e8\u7f72\u4e2d\u6700\u8017\u65f6\u7684\u73af\u8282\u3002\u6211\u4eec\u91c7\u7528\u4e24\u79cd\u4f18\u5316\u7b56\u7565：<\/p>\n
\u7b56\u75651：\u672c\u5730\u7f13\u5b58\u670d\u52a1\u5668 \u5982\u679c\u6709\u591a\u53f0\u670d\u52a1\u5668\u5728\u540c\u4e00\u4e2a\u5185\u7f51，\u53ef\u4ee5\u5148\u5728\u4e00\u53f0\u670d\u52a1\u5668\u4e0b\u8f7d\u6a21\u578b，\u7136\u540e\u901a\u8fc7\u5185\u7f51\u4f20\u8f93\u5230\u5176\u4ed6\u670d\u52a1\u5668\u3002<\/p>\n
– name: \u68c0\u67e5\u6a21\u578b\u662f\u5426\u5df2\u5b58\u5728
\n stat:
\n path: "{{ project_dir }}\/qwen\/model.safetensors"
\n register: model_file<\/p>\n
– name: \u4ece\u7f13\u5b58\u670d\u52a1\u5668\u4e0b\u8f7d\u6a21\u578b（\u5982\u679c\u5b58\u5728）
\n get_url:
\n url: "http:\/\/cache-server:8080\/models\/{{ model_id }}.tar.gz"
\n dest: "{{ project_dir }}\/model_cache.tar.gz"
\n mode: '0644'
\n when: not model_file.stat.exists and cache_server_available<\/p>\n
– name: \u89e3\u538b\u7f13\u5b58\u6a21\u578b
\n unarchive:
\n src: "{{ project_dir }}\/model_cache.tar.gz"
\n dest: "{{ project_dir }}\/qwen"
\n remote_src: yes
\n when: not model_file.stat.exists and cache_server_available<\/p>\n
– name: \u4eceModelScope\u4e0b\u8f7d\u6a21\u578b（\u5907\u7528\u65b9\u6848）
\n shell: |
\n cd {{ project_dir }}
\n python3 -c "
\n from modelscope import snapshot_download
\n snapshot_download('{{ model_id }}',
\n cache_dir='{{ project_dir }}\/qwen',
\n local_files_only=False)
\n "
\n when: not model_file.stat.exists and not cache_server_available
\n async: 3600 # \u5f02\u6b65\u6267\u884c，\u6700\u591a\u7b49\u5f851\u5c0f\u65f6
\n poll: 30 # \u6bcf30\u79d2\u68c0\u67e5\u4e00\u6b21<\/p>\n
\u7b56\u75652：\u5e76\u884c\u4e0b\u8f7d \u5bf9\u4e8e\u6ca1\u6709\u7f13\u5b58\u670d\u52a1\u5668\u7684\u60c5\u51b5，\u6211\u4eec\u53ef\u4ee5\u5728\u591a\u53f0\u670d\u52a1\u5668\u4e0a\u5e76\u884c\u4e0b\u8f7d，\u4f46\u9700\u8981\u63a7\u5236\u5e76\u53d1\u6570\u907f\u514d\u7f51\u7edc\u62e5\u5835\u3002<\/p>\n
– name: \u90e8\u7f72\u524d\u7aef\u6587\u4ef6
\n copy:
\n src: "files\/chat.html"
\n dest: "{{ project_dir }}\/chat.html"
\n mode: '0644'<\/p>\n
– name: \u90e8\u7f72\u4ee3\u7406\u670d\u52a1\u5668
\n template:
\n src: "templates\/proxy_server.py.j2"
\n dest: "{{ project_dir }}\/proxy_server.py"
\n mode: '0755'
\n vars:
\n vllm_host: "localhost"
\n vllm_port: "{{ vllm_port }}"
\n web_port: "{{ web_port }}"<\/p>\n
– name: \u90e8\u7f72\u542f\u52a8\u811a\u672c
\n copy:
\n src: "scripts\/start_all.sh"
\n dest: "{{ project_dir }}\/start_all.sh"
\n mode: '0755'<\/p>\n
– name: \u914d\u7f6eSupervisor\u670d\u52a1
\n template:
\n src: "templates\/qwen-chat.conf.j2"
\n dest: "\/etc\/supervisor\/conf.d\/qwen-chat.conf"
\n mode: '0644'<\/p>\n
– name: \u91cd\u8f7dSupervisor\u914d\u7f6e
\n shell: supervisorctl reread && supervisorctl update<\/p>\n
– name: \u542f\u52a8Qwen\u804a\u5929\u670d\u52a1
\n supervisorctl:
\n name: qwen-chat
\n state: started<\/p>\n
\u521b\u5efa\u6a21\u677f\u6587\u4ef6 templates\/proxy_server.py.j2：<\/p>\n
#!\/usr\/bin\/env python3
\nimport http.server
\nimport socketserver
\nimport requests
\nimport json
\nimport logging
\nfrom urllib.parse import urlparse<\/p>\n
# \u914d\u7f6e\u53c2\u6570
\nVLLM_HOST = "{{ vllm_host }}"
\nVLLM_PORT = {{ vllm_port }}
\nWEB_PORT = {{ web_port }}<\/p>\n
logging.basicConfig(
\n level=logging.INFO,
\n format='%(asctime)s – %(levelname)s – %(message)s'
\n)
\nlogger = logging.getLogger(__name__)<\/p>\n
class ProxyHandler(http.server.SimpleHTTPRequestHandler):
\n def do_GET(self):
\n # \u5904\u7406\u9759\u6001\u6587\u4ef6\u8bf7\u6c42
\n if self.path == '\/' or self.path == '\/chat.html':
\n self.path = '\/chat.html'
\n return super().do_GET()
\n elif self.path.endswith('.js') or self.path.endswith('.css'):
\n return super().do_GET()
\n else:
\n self.send_error(404, "File not found")<\/p>\n
def do_POST(self):
\n # \u8f6c\u53d1API\u8bf7\u6c42\u5230vLLM
\n if self.path == '\/v1\/chat\/completions':
\n content_length = int(self.headers['Content-Length'])
\n post_data = self.rfile.read(content_length)<\/p>\n
try:
\n # \u8f6c\u53d1\u8bf7\u6c42\u5230vLLM
\n vllm_url = f"http:\/\/{VLLM_HOST}:{VLLM_PORT}\/v1\/chat\/completions"
\n response = requests.post(
\n vllm_url,
\n data=post_data,
\n headers={'Content-Type': 'application\/json'},
\n timeout=30
\n )<\/p>\n
# \u8fd4\u56de\u54cd\u5e94
\n self.send_response(response.status_code)
\n self.send_header('Content-Type', 'application\/json')
\n self.send_header('Access-Control-Allow-Origin', '*')
\n self.end_headers()
\n self.wfile.write(response.content)<\/p>\n
logger.info(f"Forwarded request to vLLM, status: {response.status_code}")<\/p>\n
except Exception as e:
\n logger.error(f"Error forwarding request: {e}")
\n self.send_error(500, f"Internal Server Error: {str(e)}")
\n else:
\n self.send_error(404, "API endpoint not found")<\/p>\n
if __name__ == "__main__":
\n with socketserver.TCPServer(("", WEB_PORT), ProxyHandler) as httpd:
\n logger.info(f"Proxy server started on port {WEB_PORT}")
\n logger.info(f"vLLM backend: {VLLM_HOST}:{VLLM_PORT}")
\n httpd.serve_forever()<\/p>\n
\u51c6\u5907\u597d\u6240\u6709Playbook\u548c\u914d\u7f6e\u6587\u4ef6\u540e，\u6267\u884c\u90e8\u7f72\u547d\u4ee4：<\/p>\n
# 1. \u6d4b\u8bd5\u8fde\u63a5\u6240\u6709\u670d\u52a1\u5668
\nansible all_servers -i inventory.yml -m ping<\/p>\n
# 2. \u6267\u884c\u90e8\u7f72Playbook（\u9650\u52365\u53f0\u5e76\u53d1）
\nansible-playbook -i inventory.yml deploy_qwen.yml –limit 5 –forks 5<\/p>\n
# 3. \u67e5\u770b\u90e8\u7f72\u8fdb\u5ea6
\ntail -f \/var\/log\/ansible\/deploy.log<\/p>\n
# 4. \u68c0\u67e5\u90e8\u7f72\u7ed3\u679c
\nansible all_servers -i inventory.yml -m shell -a "supervisorctl status qwen-chat"<\/p>\n
\u5bf9\u4e8e\u5927\u89c4\u6a21\u90e8\u7f72（\u598210\u53f0\u4ee5\u4e0a），\u5efa\u8bae\u91c7\u7528\u5206\u9636\u6bb5\u7b56\u7565：<\/p>\n
# \u7b2c\u4e00\u9636\u6bb5：\u5148\u90e8\u7f722\u53f0\u4f5c\u4e3a\u6d4b\u8bd5
\nansible-playbook -i inventory.yml deploy_qwen.yml –limit "gpu-server-01,gpu-server-02"<\/p>\n
# \u7b2c\u4e8c\u9636\u6bb5：\u9a8c\u8bc1\u6d4b\u8bd5\u670d\u52a1\u5668
\nansible gpu_servers[0:1] -i inventory.yml -m uri -a "url=http:\/\/localhost:8000\/chat.html return_content=yes"<\/p>\n
# \u7b2c\u4e09\u9636\u6bb5：\u6279\u91cf\u90e8\u7f72\u5269\u4f59\u670d\u52a1\u5668（\u5206\u6279\u6b21）
\nansible-playbook -i inventory.yml deploy_qwen.yml –limit "gpu-server-03:gpu-server-05" –forks 3
\nansible-playbook -i inventory.yml deploy_qwen.yml –limit "gpu-server-06:gpu-server-10" –forks 3<\/p>\n
\u521b\u5efa\u76d1\u63a7Playbook monitor_deployment.yml：<\/p>\n
—
\n– name: \u76d1\u63a7\u90e8\u7f72\u72b6\u6001
\n hosts: all_servers
\n gather_facts: no
\n tasks:
\n – name: \u68c0\u67e5\u670d\u52a1\u72b6\u6001
\n shell: |
\n status=$(supervisorctl status qwen-chat | awk '{print $2}')
\n if [ "$status" = "RUNNING" ]; then
\n echo "OK – Service is running"
\n exit 0
\n else
\n echo "ERROR – Service status: $status"
\n exit 1
\n fi
\n register: service_status
\n ignore_errors: yes<\/p>\n
– name: \u68c0\u67e5\u7aef\u53e3\u76d1\u542c
\n wait_for:
\n port: "{{ web_port }}"
\n host: "{{ ansible_host }}"
\n timeout: 5
\n register: port_check
\n ignore_errors: yes<\/p>\n
– name: \u68c0\u67e5API\u53ef\u7528\u6027
\n uri:
\n url: "http:\/\/{{ ansible_host }}:{{ vllm_port }}\/health"
\n method: GET
\n timeout: 10
\n register: api_check
\n ignore_errors: yes<\/p>\n
– name: \u6c47\u603b\u68c0\u67e5\u7ed3\u679c
\n debug:
\n msg: |
\n \u670d\u52a1\u5668: {{ inventory_hostname }}
\n \u670d\u52a1\u72b6\u6001: {{ service_status.stdout if service_status is succeeded else 'FAILED' }}
\n \u7aef\u53e3\u76d1\u542c: {{ 'OK' if port_check is succeeded else 'FAILED' }}
\n API\u5065\u5eb7: {{ 'OK' if api_check.status == 200 else 'FAILED' }}<\/p>\n
\u6267\u884c\u76d1\u63a7：<\/p>\n
# \u5b9e\u65f6\u76d1\u63a7\u6240\u6709\u670d\u52a1\u5668\u72b6\u6001
\nwhile true; do
\n ansible-playbook -i inventory.yml monitor_deployment.yml
\n sleep 30
\ndone<\/p>\n
\u5bf9\u4e8e\u751f\u4ea7\u73af\u5883，\u6211\u4eec\u9700\u8981\u652f\u6301\u65e0\u7f1d\u66f4\u65b0：<\/p>\n
—
\n– name: \u84dd\u7eff\u90e8\u7f72\u66f4\u65b0
\n hosts: gpu_servers
\n serial: 2 # \u6bcf\u6b21\u66f4\u65b02\u53f0，\u786e\u4fdd\u670d\u52a1\u53ef\u7528\u6027
\n vars:
\n new_version: "v1.2.0"<\/p>\n
tasks:
\n – name: \u6807\u8bb0\u670d\u52a1\u5668\u4e3a\u7ef4\u62a4\u6a21\u5f0f
\n uri:
\n url: "http:\/\/{{ ansible_host }}:{{ web_port }}\/maintenance"
\n method: POST
\n body: '{"mode": "on"}'<\/p>\n
– name: \u505c\u6b62\u65e7\u7248\u672c\u670d\u52a1
\n supervisorctl:
\n name: qwen-chat
\n state: stopped<\/p>\n
– name: \u5907\u4efd\u5f53\u524d\u7248\u672c
\n archive:
\n path: "{{ project_dir }}"
\n dest: "\/backup\/qwen-{{ ansible_date_time.date }}.tar.gz"<\/p>\n
– name: \u90e8\u7f72\u65b0\u7248\u672c
\n include_tasks: deploy_qwen.yml
\n vars:
\n version: "{{ new_version }}"<\/p>\n
– name: \u542f\u52a8\u65b0\u7248\u672c\u670d\u52a1
\n supervisorctl:
\n name: qwen-chat
\n state: started<\/p>\n
– name: \u5065\u5eb7\u68c0\u67e5
\n uri:
\n url: "http:\/\/{{ ansible_host }}:{{ web_port }}\/health"
\n method: GET
\n register: health_check
\n until: health_check.status == 200
\n retries: 10
\n delay: 5<\/p>\n
– name: \u79fb\u9664\u7ef4\u62a4\u6a21\u5f0f
\n uri:
\n url: "http:\/\/{{ ansible_host }}:{{ web_port }}\/maintenance"
\n method: POST
\n body: '{"mode": "off"}'<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e\u81ea\u52a8\u8fd0\u884c\u6d4b\u8bd5\u7528\u4f8b：<\/p>\n
—
\n– name: \u81ea\u52a8\u5316\u6d4b\u8bd5\u9a8c\u8bc1
\n hosts: all_servers
\n tasks:
\n – name: \u6d4b\u8bd5\u804a\u5929\u529f\u80fd
\n uri:
\n url: "http:\/\/localhost:{{ vllm_port }}\/v1\/chat\/completions"
\n method: POST
\n body_format: json
\n body:
\n model: "{{ model_name }}"
\n messages:
\n – role: user
\n content: "\u4f60\u597d，\u8bf7\u7528\u4e00\u53e5\u8bdd\u4ecb\u7ecd\u4f60\u81ea\u5df1"
\n temperature: 0.7
\n max_tokens: 100
\n status_code: 200
\n timeout: 30
\n register: chat_test<\/p>\n
– name: \u9a8c\u8bc1\u54cd\u5e94\u5185\u5bb9
\n assert:
\n that:
\n – "'assistant' in chat_test.json.choices[0].message.role"
\n – "chat_test.json.choices[0].message.content|length > 0"
\n fail_msg: "\u804a\u5929\u529f\u80fd\u6d4b\u8bd5\u5931\u8d25"<\/p>\n
– name: \u6d4b\u8bd5\u6587\u4ef6\u670d\u52a1
\n uri:
\n url: "http:\/\/localhost:{{ web_port }}\/chat.html"
\n method: GET
\n status_code: 200
\n register: web_test<\/p>\n
– name: \u6027\u80fd\u538b\u529b\u6d4b\u8bd5（\u53ef\u9009）
\n shell: |
\n ab -n 100 -c 10 http:\/\/localhost:{{ web_port }}\/chat.html 2>\/dev\/null | grep "Requests per second"
\n when: performance_test_enabled
\n register: perf_test<\/p>\n
– name: \u751f\u6210\u6d4b\u8bd5\u62a5\u544a
\n template:
\n src: "templates\/test_report.j2"
\n dest: "\/tmp\/deployment_test_{{ inventory_hostname }}.md"<\/p>\n
\u4f7f\u7528Ansible Vault\u7ba1\u7406\u654f\u611f\u914d\u7f6e：<\/p>\n
# 1. \u521b\u5efa\u52a0\u5bc6\u7684\u914d\u7f6e\u6587\u4ef6
\nansible-vault create group_vars\/all\/vault.yml<\/p>\n
# 2. \u7f16\u8f91\u52a0\u5bc6\u6587\u4ef6（\u8f93\u5165\u5bc6\u7801）
\nansible-vault edit group_vars\/all\/vault.yml<\/p>\n
# 3. \u5728Playbook\u4e2d\u4f7f\u7528\u52a0\u5bc6\u53d8\u91cf
\n# vault.yml\u5185\u5bb9：
\nmodel_download_token: "your-secret-token"
\napi_keys:
\n openai: "sk-xxx"
\n anthropic: "claude-xxx"<\/p>\n
# 4. \u6267\u884c\u65f6\u63d0\u4f9b\u5bc6\u7801
\nansible-playbook -i inventory.yml deploy_qwen.yml –ask-vault-pass<\/p>\n
\u95ee\u98981：\u6a21\u578b\u4e0b\u8f7d\u5931\u8d25<\/p>\n
# \u68c0\u67e5\u7f51\u7edc\u8fde\u63a5
\nansible all_servers -i inventory.yml -m shell -a "curl -I https:\/\/modelscope.cn"<\/p>\n
# \u624b\u52a8\u4e0b\u8f7d\u6a21\u578b\u5230\u7f13\u5b58\u670d\u52a1\u5668
\nansible cache_server -i inventory.yml -m shell -a "cd \/var\/cache\/models && wget https:\/\/modelscope.cn\/models\/{{ model_id }}\/repo?Revision=master"<\/p>\n
# \u4ece\u7f13\u5b58\u670d\u52a1\u5668\u5206\u53d1
\nansible gpu_servers -i inventory.yml -m synchronize -a "src=\/var\/cache\/models\/ dest={{ project_dir }}\/qwen\/"<\/p>\n
\u95ee\u98982：GPU\u5185\u5b58\u4e0d\u8db3<\/p>\n
# \u8c03\u6574vLLM\u5185\u5b58\u914d\u7f6e
\n– name: \u8c03\u6574GPU\u5185\u5b58\u4f7f\u7528\u7387
\n lineinfile:
\n path: "{{ project_dir }}\/start_all.sh"
\n regexp: '^.*–gpu-memory-utilization.*$'
\n line: ' –gpu-memory-utilization 0.5'<\/p>\n
# \u6216\u8005\u4f7f\u7528CPU\u6a21\u5f0f
\n– name: \u4f7f\u7528CPU\u6a21\u5f0f（\u5907\u7528\u65b9\u6848）
\n lineinfile:
\n path: "{{ project_dir }}\/start_all.sh"
\n regexp: '^vllm serve'
\n line: 'vllm serve "$ACTUAL_MODEL_PATH" –device cpu'
\n when: gpu_memory < '8GB'<\/p>\n
\u95ee\u98983：\u670d\u52a1\u542f\u52a8\u8d85\u65f6<\/p>\n
# \u589e\u52a0\u542f\u52a8\u8d85\u65f6\u65f6\u95f4
\nansible all_servers -i inventory.yml -m shell -a "sed -i 's\/startsecs=10\/startsecs=30\/' \/etc\/supervisor\/conf.d\/qwen-chat.conf && supervisorctl update"<\/p>\n
# \u67e5\u770b\u8be6\u7ec6\u65e5\u5fd7
\nansible all_servers -i inventory.yml -m shell -a "tail -100 {{ project_dir }}\/vllm.log"<\/p>\n
\u521b\u5efa\u76d1\u63a7\u811a\u672c monitor_services.sh：<\/p>\n
#!\/bin\/bash
\n# \u76d1\u63a7\u6240\u6709\u670d\u52a1\u5668\u7684\u670d\u52a1\u72b6\u6001<\/p>\n
INVENTORY="inventory.yml"
\nLOG_FILE="\/var\/log\/qwen_monitor.log"<\/p>\n
# \u68c0\u67e5\u5355\u53f0\u670d\u52a1\u5668
\ncheck_server() {
\n local host=$1
\n local ip=$2<\/p>\n
# \u68c0\u67e5\u670d\u52a1\u8fdb\u7a0b
\n ssh $ip "supervisorctl status qwen-chat" | grep -q "RUNNING"
\n if [ $? -ne 0 ]; then
\n echo "$(date) – ERROR: Service not running on $host" >> $LOG_FILE
\n return 1
\n fi<\/p>\n
# \u68c0\u67e5\u7aef\u53e3
\n nc -z $ip 8000
\n if [ $? -ne 0 ]; then
\n echo "$(date) – ERROR: Port 8000 not listening on $host" >> $LOG_FILE
\n return 1
\n fi<\/p>\n
# \u68c0\u67e5API
\n curl -s http:\/\/$ip:3001\/health | grep -q "healthy"
\n if [ $? -ne 0 ]; then
\n echo "$(date) – ERROR: API not healthy on $host" >> $LOG_FILE
\n return 1
\n fi<\/p>\n
echo "$(date) – OK: $host is healthy" >> $LOG_FILE
\n return 0
\n}<\/p>\n
# \u4eceAnsible\u6e05\u5355\u8bfb\u53d6\u670d\u52a1\u5668\u5217\u8868
\nservers=$(ansible-inventory -i $INVENTORY –list | jq -r '.all_servers.hosts[]')<\/p>\n
for server in $servers; do
\n ip=$(ansible-inventory -i $INVENTORY –host $server | jq -r '.ansible_host')
\n check_server $server $ip &
\ndone<\/p>\n
wait<\/p>\n
\u521b\u5efa\u7ef4\u62a4Playbook maintenance.yml：<\/p>\n
—
\n– name: \u5b9a\u671f\u7ef4\u62a4\u4efb\u52a1
\n hosts: all_servers
\n tasks:
\n – name: \u6e05\u7406\u65e5\u5fd7\u6587\u4ef6
\n shell: |
\n find {{ project_dir }} -name "*.log" -type f -size +100M -exec truncate -s 50M {} \\\\;<\/p>\n
– name: \u68c0\u67e5\u78c1\u76d8\u7a7a\u95f4
\n shell: df -h \/ | awk 'NR==2 {print $5}' | sed 's\/%\/\/'
\n register: disk_usage<\/p>\n
– name: \u78c1\u76d8\u7a7a\u95f4\u544a\u8b66
\n debug:
\n msg: "\u8b66\u544a：{{ inventory_hostname }} \u78c1\u76d8\u4f7f\u7528\u7387 {{ disk_usage.stdout }}%"
\n when: disk_usage.stdout|int > 80<\/p>\n
– name: \u66f4\u65b0\u7cfb\u7edf\u5305
\n apt:
\n upgrade: yes
\n update_cache: yes
\n when: ansible_os_family == "Debian"<\/p>\n
– name: \u91cd\u542f\u670d\u52a1（\u6bcf\u5468\u4e00\u6b21）
\n shell: supervisorctl restart qwen-chat
\n when: ansible_date_time.weekday == "Monday"<\/p>\n
\u901a\u8fc7\u8fd9\u5957Ansible\u81ea\u52a8\u5316\u90e8\u7f72\u65b9\u6848，\u6211\u4eec\u6210\u529f\u89e3\u51b3\u4e86\u5728\u591a\u53f0\u670d\u52a1\u5668\u4e0a\u6279\u91cf\u90e8\u7f72Qwen3-VL-8B AI\u804a\u5929\u7cfb\u7edf\u7684\u96be\u9898\u3002\u8ba9\u6211\u4eec\u56de\u987e\u4e00\u4e0b\u5173\u952e\u8981\u70b9：<\/p>\n
\u90e8\u7f72\u6548\u7387\u5927\u5e45\u63d0\u5347<\/p>\n
\u914d\u7f6e\u7ba1\u7406\u66f4\u52a0\u89c4\u8303<\/p>\n
\u8fd0\u7ef4\u76d1\u63a7\u5168\u9762\u8986\u76d6<\/p>\n
\u6269\u5c55\u6027\u826f\u597d<\/p>\n
\u5b9e\u9645\u90e8\u7f72\u5efa\u8bae<\/p>\n
\u8fd9\u5957\u65b9\u6848\u4e0d\u4ec5\u9002\u7528\u4e8eQwen3-VL-8B，\u7ecf\u8fc7\u9002\u5f53\u8c03\u6574\u540e，\u4e5f\u53ef\u4ee5\u7528\u4e8e\u90e8\u7f72\u5176\u4ed6AI\u6a21\u578b\u548c\u5e94\u7528\u3002\u81ea\u52a8\u5316\u90e8\u7f72\u662fAI\u5de5\u7a0b\u5316\u7684\u91cd\u8981\u4e00\u73af，\u80fd\u663e\u8457\u63d0\u5347\u56e2\u961f\u7684\u751f\u4ea7\u529b\u548c\u7cfb\u7edf\u7684\u53ef\u9760\u6027\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Qwen3-VL-8B\u5f00\u6e90\u90e8\u7f72\u6559\u7a0b：Ansible\u81ea\u52a8\u5316\u811a\u672c\u5b9e\u73b010\u53f0\u670d\u52a1\u5668\u6279\u91cf\u90e8\u7f72
\n1. \u5f15\u8a00
\n\u60f3\u8c61\u4e00\u4e0b，\u4f60\u9700\u8981\u572810\u53f0\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u4e00\u4e2a\u5b8c\u6574\u7684AI\u804a\u5929\u7cfb\u7edf\u3002\u6bcf\u53f0\u670d\u52a1\u5668\u90fd\u8981\u5b89\u88c5Python\u73af\u5883\u3001\u4e0b\u8f7d\u51e0\u5341GB\u7684\u6a21\u578b\u6587\u4ef6\u3001\u914d\u7f6evLLM\u63a8\u7406\u5f15\u64ce\u3001\u8bbe\u7f6e\u53cd\u5411\u4ee3\u7406\u3001\u90e8\u7f72\u524d\u7aef\u754c\u9762\u2026\u2026\u5982\u679c\u624b\u52a8\u64cd\u4f5c，\u4e00\u53f0\u670d\u52a1\u5668\u53ef\u80fd\u5c31\u8981\u6298\u817e\u534a\u5929，10\u53f0\u670d\u52a1\u5668\u7b80\u76f4\u662f\u4e2a\u5669\u68a6\u3002
\n\u8fd9\u5c31\u662f\u6211\u4eec\u4eca\u5929\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002\u6211\u5c06\u5206\u4eab\u4e00\u4e2a\u57fa\u4e8eAnsible\u7684\u81ea\u52a8\u5316\u90e8\u7f72\u65b9\u6848&#<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[7591,1064,956,118],"topic":[],"class_list":["post-87877","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-ansible","tag-956","tag-118"],"yoast_head":"\n