# CI service. GPU and CPU builds run in parallel and must not share a tag:
# each one deletes its own tag before building. Bind-mount the checkout onto
# itself for pytest, and also onto /app. The image editable-installs
# agentinfer there, and the vllm console script loads that path rather than
# the pytest working directory.
services:
  agentinfer-ci:
    build:
      context: ..
      dockerfile: docker/Dockerfile
    image: agentinfer-ci:${BUILDKITE_BUILD_ID:-local}-gpu
    # Runs as root, then ci-entrypoint.sh switches to the UID that owns the
    # checkout (the working directory). Do not set `user:` here.
    entrypoint: ["/usr/local/bin/ci-entrypoint.sh"]
    working_dir: ${BUILDKITE_BUILD_CHECKOUT_PATH:-/app}
    network_mode: host
    ipc: host
    shm_size: 64gb
    cap_add: [SYS_PTRACE]
    volumes:
      - /home/models:/home/models
      - ${BUILDKITE_BUILD_CHECKOUT_PATH:-/app}:${BUILDKITE_BUILD_CHECKOUT_PATH:-/app}
      - ${BUILDKITE_BUILD_CHECKOUT_PATH:-/app}:/app
    environment:
      HF_ENDPOINT: https://hf-mirror.com
      HF_HOME: /home/models
      HF_HUB_CACHE: /home/models/hub
      HOME: /tmp/ci-home
      PYTEST_ADDOPTS: -o cache_dir=/tmp/pytest-cache
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # Same Dockerfile as agentinfer-ci, different tag. !override clears the GPU
  # reservation so this service can start on a machine with no NVIDIA runtime.
  # An empty mapping would be merged and the GPU devices would remain.
  agentinfer-cpu:
    extends:
      service: agentinfer-ci
    image: agentinfer-ci:${BUILDKITE_BUILD_ID:-local}-cpu
    deploy: !override {}