APPs/llama-swap/docker-compose.yaml

#version: '3.8'

services:
  llama-swap:
    container_name: llamaswap
    image: ghcr.io/mostlygeek/llama-swap:vulkan
    restart: unless-stopped
    pull_policy: always
    network_mode: host
    privileged: true
    shm_size: '32gb'
    volumes:
      # Novo mapeamento unificado de cache (compatível com HF migration)
      - /home/lvivas/Modelos/models:/root/.cache
      # Acesso às gramáticas para o orquestrador validar os caminhos
      - /home/lvivas/Modelos/grammar:/tmp/grammar
      # Configurações do projeto
      - ./config:/config
      # Comunicação com o Docker Host (indispensável)
      - /var/run/docker.sock:/var/run/docker.sock
      - /usr/bin/docker:/usr/bin/docker
      # Cache de contexto compartilhado
      - /tmp/cache_file:/tmp/cache_file
    environment:
      # Opcional: Garante que o orquestrador saiba onde procurar as coisas
      - XDG_CACHE_HOME=/root/.cache
    devices:
      - /dev/dri:/dev/dri
      - /dev/accel/accel0:/dev/accel/accel0
    group_add:
      - "44"
      - "991"
    cap_add:
      - SYS_ADMIN
      - SYS_RAWIO
      - IPC_LOCK
      - SYS_RESOURCE
    ulimits:
      memlock: -1
      stack: 67108864
    entrypoint: /app/llama-swap -config /config/config.yaml --listen :65000
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:65000/"]
      interval: 30s
      timeout: 10s
      retries: 3

  # ESTE BLOCO DEVE ESTAR ALINHADO COM O llama-swap:
  watch-config:
    image: docker:latest
    container_name: llamaswap-watcher
    restart: unless-stopped
    volumes:
      - ./config:/config
      - /home/lvivas/Modelos/models:/root/.cache
      - /var/run/docker.sock:/var/run/docker.sock
      - /tmp/cache_file:/tmp/cache_file
    # O entrypoint garante que o script tenha permissão antes de rodar
    entrypoint: ["sh", "-c", "chmod +x /config/watch-config.sh && /config/watch-config.sh"]