multiprocessing — 基于进程的并行：https://docs.python.org/zh-cn/3.11/library/multiprocessing.html
import sys
from fastchat.serve.controller import Controller
from fastchat.serve.model_worker import ModelWorker
from fastchat.serve.openai_api_server import app, CORSMiddleware, app_settingsimport sys
import uuid
import uvicorn
from fastapi import FastAPI# 创建不同的 FastAPI 应用实例
controller_app = FastAPI(title="FastChat Controller Service")
worker_app = FastAPI(title="FastChat Model Worker Service")
api_app = FastAPI(title="FastChat OpenAI API Service")def start_controller():"""启动 FastChat Controller 服务分布式系统设计中常见的一种优化策略，决定如何分配任务给不同的model_worker（或服务器）- LOTTERY：这种方法系统会随机选择一个worker。不考虑worker的当前负载或任何其他因素。- SHORTEST_QUEUE：这种方法会选择当前队列长度最短的worker，也就是当前负载最小的工人。"""controller = Controller(dispatch_method="shortest_queue")# sys.modules 是一个字典，它存储了已经加载的模块。每个键是一个模块名，每个值是一个模块对象。# 这种机制使得Python在导入模块时可以检查模块是否已经在sys.modules中，如果是，就直接使用已经加载的模块，避免重复加载。sys.modules["fastchat.serve.controller"].controller = controllercontroller_app.title = "FastChat Controller"controller_app._controller = controlleruvicorn.run(controller_app, host="192.168.110.131", port=21001)def start_model_worker():"""启动 Model Worker 服务"""worker_id = str(uuid.uuid4())[:8]worker = ModelWorker(controller_addr="http://192.168.110.131:21001",worker_addr="http://192.168.110.131:21002",worker_id=worker_id,limit_worker_concurrency=5,no_register=False,# no_register=True,model_path="/home/00_rag/model/ZhipuAI/chatglm3-6b",num_gpus=4,model_names=["chatglm3-6b"],device="cuda",max_gpu_memory="22GiB",)worker_app.title = f"FastChat LLM Server ChaGLM3-6b"worker_app._worker = workeruvicorn.run(worker_app, host="192.168.110.131", port=21002)def start_openai_api_server():"""启动 OpenAI API 服务"""api_app.add_middleware(CORSMiddleware,allow_credentials=True, # 允许前端请求携带认证信息（如 cookiesallow_origins=["*"], # 允许所有域名的请求，星号表示不限制任何域。allow_methods=["*"], # 允许所有的 HTTP 方法。allow_headers=["*"], # 允许所有的 HTTP 头)app_settings.controller_address = "http://192.168.110.131:21001"app_settings.api_keys = []api_app.title = "FastChat OpenAI API Server"uvicorn.run(api_app, host="192.168.110.131", port=8000)from multiprocessing import Process
import timedef start_services_in_processes():# 创建进程# controller_process = Process(target=start_controller)worker_process = Process(target=start_model_worker)api_server_process = Process(target=start_openai_api_server)# 启动进程# controller_process.start()worker_process.start()api_server_process.start()# 等待所有进程完成# controller_process.join()worker_process.join()api_server_process.join()if __name__ == '__main__':start_services_in_processes()
