docker-compose.test.yml

# inferhost v0.5+ test stack.
# Use: ./run.sh docker-build, then docker-smoke / docker-test / docker-shell.
# Requires NVIDIA Container Toolkit on the host for --gpus all.

services:
  inferhost:
    build:
      context: .
      dockerfile: Dockerfile
    image: inferhost-test:v0.5
    container_name: inferhost-test
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITIES: compute,utility
      INFERHOST_GATEWAY_PORT: "9001"
      # llama-swap stays loopback (v0.5 default); inside the container we
      # don't need to reach it externally — only LiteLLM is exposed.
      INFERHOST_SWAP_PORT: "9090"
      # Force the prebuilt llama-server target if needed (uncomment for
      # CPU-only host builds). The default uses the binaries.py auto-detect.
      # INFERHOST_LLAMACPP_BACKEND: cuda
    ports:
      # Published on a non-default host port: the host's own inferhost
      # daemons (llama-swap :9090, litellm :9001, tts :9096) may already be
      # bound to the defaults, and `docker compose up` would fail to start
      # if we tried to reuse them. Container-internal port stays 9001
      # (INFERHOST_GATEWAY_PORT below) — docker-functional talks to it via
      # `docker compose exec`, which never touches this host mapping.
      - "19001:9001"  # LiteLLM gateway (host-side only; avoids host's live :9001)
    volumes:
      - hf-cache:/inferhost/hf-cache
      - inferhost-config:/inferhost/config
      - inferhost-data:/inferhost/data
    stdin_open: true
    tty: true
    # Keep the container alive so `docker compose run` / `exec` are cheap.
    command: ["sleep", "infinity"]

volumes:
  hf-cache:
  inferhost-config:
  inferhost-data: