# docker-compose for local inferscope runs
#
# Builds the image from the repository's Dockerfile and runs inferscope
# with GPU access. Adjust the `command` block to your engine endpoint
# and model. Output is written to ./output/ on the host.
#
# Quick start:
#   docker compose -f deploy/docker-compose.yml up --build
#
# To inspect the JSON report after the run:
#   cat output/report.json | jq .
#
# Requires:
#   - Docker 24+ (for compose v2 syntax)
#   - NVIDIA Container Toolkit on the host
#   - An OpenAI-compatible engine reachable from the container

services:
  inferscope:
    build:
      context: ..
      dockerfile: Dockerfile
    image: inferscope:local
    container_name: inferscope-run

    # GPU passthrough — requires NVIDIA Container Toolkit installed on host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

    # Replace these flags with the ones for your actual run
    command:
      - --endpoint
      - http://host.docker.internal:8080
      - --model
      - your-model-name
      - --prompt
      - "Hello, world"
      - --gpu

    # Mount host directory for JSON output (if using --json flag)
    volumes:
      - ./output:/home/inferscope/output

    # If the engine runs on the same host, expose it through the host network
    # extra_hosts:
    #   - "host.docker.internal:host-gateway"