# ODS — Apple Silicon Overlay (macOS / Metal)
# Uses llama.cpp server image built for Apple Metal acceleration.
# On macOS, Docker runs in a Linux VM — GPU passthrough to Metal is NOT
# natively supported. This overlay uses a CPU-optimized llama.cpp build.
# For full Metal acceleration, run llama-server on the host (outside Docker)
# and point OLLAMA_PORT at it.
#
# Use with: docker compose -f docker-compose.base.yml -f docker-compose.apple.yml up -d

services:
  llama-server:
    image: ghcr.io/ggml-org/llama.cpp:server-b8248
    platform: linux/arm64
    deploy:
      resources:
        limits:
          cpus: '${LLAMA_CPU_LIMIT:-8.0}'
          memory: ${LLAMA_SERVER_MEMORY_LIMIT:-32G}
        reservations:
          cpus: '${LLAMA_CPU_RESERVATION:-2.0}'
          memory: 4G
    environment:
      # Hint to llama.cpp for ARM NEON optimizations
      - LLAMA_NO_METAL=1