docker-compose.cloud.yaml

# =============================================================================
# Hive — Cloud GPU deployment override
#
# Usage:
#   docker compose -f docker-compose.yaml -f docker-compose.cloud.yaml up -d
#
# Target: Linux cloud VMs with NVIDIA GPUs
#   AWS g5/p4, GCP a2/a3, Lambda Labs, CoreWeave, Linux Desktop instances
#
# This file overrides only the fields that differ from the base compose file.
# All services, volumes, and networks are inherited from docker-compose.yaml.
#
# Requirements:
#   - NVIDIA Container Toolkit installed and configured
#   - docker info | grep -i runtime  →  should show "nvidia"
#   - For rootless Docker: nvidia-ctk runtime configure --runtime=docker
#       --config=$HOME/.config/docker/daemon.json && systemctl --user restart docker
# =============================================================================

name: hive

services:

  # ---------------------------------------------------------------------------
  # hive-api — bind to all interfaces so cloud load balancers can reach it
  # ---------------------------------------------------------------------------
  hive-api:
    ports:
      # Cloud: expose on all interfaces (sit behind nginx/caddy or security group)
      - "0.0.0.0:8000:8000"
    environment:
      HIVE_DATA_DIR: /data
      PYTHONUNBUFFERED: "1"
      OLLAMA_HOST: "${OLLAMA_HOST:-http://ollama:11434}"
      # Raise the API worker count for multi-core cloud VMs
      WEB_CONCURRENCY: "${WEB_CONCURRENCY:-4}"
    deploy:
      resources:
        limits:
          # Cloud VMs typically have more RAM; raise ceiling
          memory: 4G
        reservations:
          memory: 512M
    healthcheck:
      # Faster health checks — cloud orchestrators need quick convergence
      interval: 15s
      timeout: 5s
      start_period: 20s
      retries: 3

  # ---------------------------------------------------------------------------
  # hive-web — bind to all interfaces for external access
  # ---------------------------------------------------------------------------
  hive-web:
    ports:
      - "0.0.0.0:3000:3000"
    environment:
      # On cloud instances the public IP is the ingress; users set this in .env
      NEXT_PUBLIC_API_URL: "${NEXT_PUBLIC_API_URL:-http://localhost:8000}"
      NODE_ENV: production
    deploy:
      resources:
        limits:
          memory: 1G
        reservations:
          memory: 256M
    healthcheck:
      interval: 15s
      timeout: 5s
      start_period: 30s
      retries: 3

  # ---------------------------------------------------------------------------
  # ollama — full GPU passthrough with cloud-tuned environment
  # ---------------------------------------------------------------------------
  ollama:
    ports:
      # Bind to loopback only; external access goes through the API
      - "127.0.0.1:11434:11434"

    # Explicit NVIDIA runtime declaration — required on some cloud providers
    # (supplements the deploy.resources.reservations block in the base file)
    runtime: nvidia

    environment:
      OLLAMA_HOST: "0.0.0.0:11434"
      OLLAMA_KEEP_ALIVE: "${OLLAMA_KEEP_ALIVE:-30m}"

      # Flash Attention significantly reduces VRAM usage on Ampere/Ada/Hopper GPUs
      OLLAMA_FLASH_ATTENTION: "${OLLAMA_FLASH_ATTENTION:-1}"

      # Allow more parallel requests — cloud VMs have more RAM and bandwidth
      OLLAMA_NUM_PARALLEL: "${OLLAMA_NUM_PARALLEL:-4}"

      # Pass-through NVIDIA runtime variables required by the NVIDIA Container Toolkit
      NVIDIA_VISIBLE_DEVICES: "${NVIDIA_VISIBLE_DEVICES:-all}"
      NVIDIA_DRIVER_CAPABILITIES: "${NVIDIA_DRIVER_CAPABILITIES:-compute,utility}"

      # Optional: set a VRAM ceiling so Ollama doesn't OOM other processes.
      # Expressed as a fraction of total VRAM (0.9 = 90%).
      # OLLAMA_GPU_OVERHEAD: "0"

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
        limits:
          # Prevent Ollama's CPU threads from monopolising the VM
          # GPU memory is managed by CUDA — this only limits host RAM
          memory: "${OLLAMA_MEMORY_LIMIT:-16G}"

    healthcheck:
      # Faster convergence on cloud — models load quickly from NVMe
      interval: 10s
      timeout: 10s
      start_period: 20s
      retries: 6

  # ---------------------------------------------------------------------------
  # ollama-init — pull the right model based on GPU VRAM
  #
  # Override OLLAMA_PULL_MODEL in your .env to choose a different model:
  #   OLLAMA_PULL_MODEL=nemotron          # 70B — needs 48 GB+ VRAM (A100, H100)
  #   OLLAMA_PULL_MODEL=nemotron-small    # 4B  — runs on 8 GB+ VRAM (default)
  #   OLLAMA_PULL_MODEL=llama3.3:70b      # Meta Llama 3.3 70B
  # ---------------------------------------------------------------------------
  ollama-init:
    environment:
      OLLAMA_HOST: "http://ollama:11434"
    command:
      - |
        MODEL="${OLLAMA_PULL_MODEL:-nemotron-small}"
        echo "Hive cloud init: checking for model: $MODEL ..."
        if ollama list 2>/dev/null | grep -qF "$MODEL"; then
          echo "$MODEL already present — skipping pull."
        else
          echo "Pulling $MODEL ..."
          ollama pull "$MODEL" && echo "Pull complete." \
            || echo "Pull failed — retry with: docker compose exec ollama ollama pull $MODEL"
        fi

  # ---------------------------------------------------------------------------
  # nginx — HTTPS reverse proxy (optional; activate with --profile https)
  #
  # Usage:
  #   docker compose -f docker-compose.yaml -f docker-compose.cloud.yaml \
  #     --profile https up -d
  #
  # Place your TLS certificate files at:
  #   ./nginx/certs/hive.crt
  #   ./nginx/certs/hive.key
  # Or let Caddy handle ACME automatically (see caddy service below).
  # ---------------------------------------------------------------------------
  nginx:
    image: nginx:1.27-alpine
    container_name: hive-nginx
    profiles: ["https"]
    restart: unless-stopped

    ports:
      - "0.0.0.0:80:80"
      - "0.0.0.0:443:443"

    volumes:
      - ./nginx/nginx.conf:/etc/nginx/nginx.conf:ro
      - ./nginx/certs:/etc/nginx/certs:ro

    networks:
      - council-net

    depends_on:
      hive-api:
        condition: service_healthy
      hive-web:
        condition: service_healthy

    healthcheck:
      test: ["CMD", "nginx", "-t"]
      interval: 30s
      timeout: 5s
      retries: 3

  # ---------------------------------------------------------------------------
  # caddy — automatic HTTPS via ACME/Let's Encrypt (optional profile: caddy)
  #
  # Usage:
  #   HIVE_DOMAIN=hive.example.com \
  #   docker compose -f docker-compose.yaml -f docker-compose.cloud.yaml \
  #     --profile caddy up -d
  # ---------------------------------------------------------------------------
  caddy:
    image: caddy:2-alpine
    container_name: hive-caddy
    profiles: ["caddy"]
    restart: unless-stopped

    ports:
      - "0.0.0.0:80:80"
      - "0.0.0.0:443:443"
      - "0.0.0.0:443:443/udp"  # HTTP/3

    environment:
      HIVE_DOMAIN: "${HIVE_DOMAIN:-localhost}"

    volumes:
      - ./caddy/Caddyfile:/etc/caddy/Caddyfile:ro
      - caddy-data:/data
      - caddy-config:/config

    networks:
      - council-net

    depends_on:
      hive-api:
        condition: service_healthy
      hive-web:
        condition: service_healthy

    healthcheck:
      test: ["CMD", "caddy", "version"]
      interval: 30s
      timeout: 5s
      retries: 3

# ---------------------------------------------------------------------------
# Extra volumes for optional proxy services
# ---------------------------------------------------------------------------
volumes:
  hive-data:
  ollama-models:
  caddy-data:
  caddy-config: