docker-compose.yml

version: "3.3"

services:
  triton:
    build:
      context: .
      dockerfile: Dockerfile
      args:
        DGPU_ARCHS: 75
        TRITON_VERSION: "21.10"
        TRT_OSS_CHECKOUT_TAG: release/8.2
        TENSORRT_REPO: https://github.com/hiennguyen9874/TensorRT
    command:
      [
        "tritonserver",
        "--model-repository=/models",
        "--strict-model-config=false",
        "--log-verbose=1",
        "--backend-config=python,shm-default-byte-size=268435456",
      ]
    volumes:
      - ./models:/models
    environment:
      # USE_NEW_NVSTREAMMUX: "yes"
      NVIDIA_VISIBLE_DEVICES: 0
      NVIDIA_DRIVER_CAPABILITIES: all
    expose:
      - 8000
      - 8001
      - 8002
    ports:
      - 48000:8000
      - 48001:8001
      - 48002:8002
    runtime: nvidia
    restart: unless-stopped
    ipc: host
    shm_size: 4g
    ulimits:
      memlock: -1
      stack: 67108864
    stdin_open: true # docker run -i
    tty: true # docker run -t
    # https://github.com/NVIDIA/nccl/issues/360#issuecomment-670650867
    pid: host