source ～/vllm-env/bin/activate # 基础依赖 pip install modelscope vllmimport os from modelscope import snapshot_download # ====================== # 配置模型 # ====================== LLM_MODEL = "Qwen/Qwen2.5-0.5B" VLM_MODEL = "OpenBMB/MiniCPM-V-2_6-int4" # 或 MiniCPM-V-2_6（非量化） LLM_DIR = f"./models/{LLM_MODEL.replace('/', '_')}" VLM_DIR = f"./models/{VLM_MODEL.replace('/', '_')}" # ====================== # 下载模型（ModelScope） # ====================== for name, path in [(LLM_MODEL, LLM_DIR), (VLM_MODEL, VLM_DIR)]: if not os.path.exists(path) or not os.listdir(path): print(f"📥 下载模型: {name}") os.makedirs(path, exist_ok=True) snapshot_download(model_id=name, local_dir=path) else: print(f"✅ 模型已存在: {path}") # ====================== # 启动服务 # ====================== print("\n🚀 启动 LLM 服务 (端口 8000)...") llm_cmd = ( f"python -m vllm.entrypoints.openai.api_server " f"--model {LLM_DIR} " f"--host 0.0.0.0 --port 8000 " f"--dtype bfloat16" ) print("🚀 启动 VLM 服务 (端口 8001)...") vlm_cmd = ( f"python -m vllm.entrypoints.openai.api_server " f"--model {VLM_DIR} " f"--host 0.0.0.0 --port 8001 " f"--dtype bfloat16 " f"--trust-remote-code" # ⚠️ 关键参数！ ) print("\n🌐 访问地址:") print(f" - LLM: http://localhost:8000/v1") print(f" - VLM: http://localhost:8001/v1") # 后台启动 VLM，前台运行 LLM os.system(f"nohup {vlm_cmd} > vlm.log 2>&1 &") os.system(llm_cmd)
)
