services:
  llama-vulkan:
    image: ghcr.io/ggml-org/llama.cpp:server-vulkan
    container_name: llama-server-vulkan
    ports:
      - "${HOST_PORT:-8080}:8080"
    volumes:
      - /home/andy/.lmstudio/models/lmstudio-community:/root/.cache/llama.cpp/
      # Optionally mount a local .gguf file directly:
      # - /absolute/path/to/your-model.gguf:/model.gguf:ro
    devices:
      - /dev/dri:/dev/dri
      - /dev/kfd:/dev/kfd
    environment:
      # Uncomment to select specific GPU(s) if you have multiple
      # GGML_VK_VISIBLE_DEVICES: "0"  # Use first Vulkan device
      
      # Model loading: Option A - Download from Hugging Face (recommended)
      LLAMA_ARG_HF_REPO: "${HF_REPO:-ggml-org/Qwen3-1.7B-GGUF}"
      
      # Model loading: Option B - Use local file (uncomment, comment HF_REPO)
      # LLAMA_ARG_MODEL: "/model.gguf"
      
      # Server settings
      LLAMA_ARG_HOST: "0.0.0.0"
      LLAMA_ARG_PORT: "8878"
      LLAMA_ARG_N_GPU_LAYERS: "99"
      LLAMA_ARG_CTX_SIZE: "${CTX_SIZE:-8000}"
      
      # Optional performance settings
      LLAMA_ARG_BATCH_SIZE: "${BATCH_SIZE:-512}"
      LLAMA_ARG_THREADS: "${THREADS:-8}"
      
    restart: unless-stopped
    command: [
      "--jinja",  # Enable Jinja templates for chat formatting
      "--flash-attn"  # Enable flash attention if supported
    ]