Пример того, как в докеркомпозе можно запускать VLLM

services:

  vllm:
    image: "vllm/vllm-openai:v0.8.5.post1"
    container_name: "vllm-qwen3"
    runtime: nvidia

    ulimits:
      nofile: 65535

    deploy:
      resources:
        reservations:
          devices:
            - capabilities: ["gpu"]

    volumes:
      - ./cache:/root/.cache/huggingface

    environment:
      - VLLM_USE_V1=0
      - VLLM_ATTENTION_BACKEND="FLASH_ATTN"
      #- HUGGING_FACE_HUB_TOKEN=<secret>

    ports:
      - "12900:8000"

    ipc: host

    entrypoint: "vllm serve Qwen/Qwen3-8B --host 0.0.0.0 --max-model-len 2K --enable-reasoning --reasoning-parser deepseek_r1"