# One container: llama.cpp on :8010 and Jev's decision contract on :8011.
#
# Bonsai is small enough to share the box, but the Spark's unified memory means
# an overshoot hangs the host, so the entrypoint refuses to start without
# headroom. Pick the quantisation with MODEL_FILE: the 1-bit build is faster and
# fine for classification, the ternary build is the one that does arithmetic.
services:
  bjev:
    build: .
    image: bjev-spark:latest
    container_name: bjev
    network_mode: host
    runtime: nvidia
    ipc: host
    restart: unless-stopped
    ulimits:
      core: 0
    volumes:
      - ${MODELS_DIR:-./models}:/models:ro
    environment:
      - MODEL=/models/${MODEL_NAME:-bonsai-ternary-27b}/${MODEL_FILE:-Ternary-Bonsai-27B-PQ2_0.gguf}
      - DRAFT=${DRAFT_FILE:-}
      - SERVED_NAME=${SERVED_NAME:-bonsai}
      - PORT=${PORT:-8010}
      - DECISION_PORT=${DECISION_PORT:-8011}
      - CTX=${CTX:-8192}
      - NGL=${NGL:-99}
      - PARALLEL=${PARALLEL:-1}
      - HEADROOM_GB=${HEADROOM_GB:-8}
      - API_KEY=${API_KEY:-}
      - EXTRA_ARGS=${EXTRA_ARGS:-}
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility