# Docker Compose DGX Overlay
# Platform-specific overrides for NVIDIA DGX (Grace ARM + Blackwell GPU)
#
# Usage: docker compose -f docker-compose.yml -f docker-compose.dgx.yml up -d

volumes:
  ollama_models:
    driver: local

services:
  # Control Panel Backend - DGX Configuration
  control-panel-backend:
    extra_hosts:
      - "host.docker.internal:host-gateway"
      - "ollama-host:host-gateway"
    healthcheck:
      start_period: 120s

  # Tenant Backend - DGX Environment
  tenant-backend:
    extra_hosts:
      - "host.docker.internal:host-gateway"
      - "ollama-host:host-gateway"
    environment:
      ENVIRONMENT: production
      DEBUG: "false"
      # Linux Docker networking for embedding service
      EMBEDDING_ENDPOINT: http://vllm-embeddings:8000
    healthcheck:
      start_period: 120s

  # Tenant PostgreSQL Primary - DGX Performance Tuning
  tenant-postgres-primary:
    environment:
      # DGX Performance settings - 128GB memory optimized
      POSTGRES_SHARED_BUFFERS: 4GB
      POSTGRES_EFFECTIVE_CACHE_SIZE: 96GB
      POSTGRES_MAINTENANCE_WORK_MEM: 1GB
      POSTGRES_MAX_CONNECTIONS: 500
      POSTGRES_WORK_MEM: 256MB
    deploy:
      resources:
        limits:
          memory: 24G
        reservations:
          memory: 16G
    labels:
      - "gt2.platform=dgx"
      - "gt2.architecture=grace-arm"

  # Resource Cluster - DGX Host Access
  resource-cluster:
    extra_hosts:
      - "host.docker.internal:host-gateway"
      - "ollama-host:host-gateway"
    environment:
      ENVIRONMENT: production
      DEBUG: "false"

  # VLLM Embeddings Service - DGX Grace ARM Optimized
  vllm-embeddings:
    platform: linux/arm64
    build:
      context: .
      dockerfile: .deployment/docker/Dockerfile.vllm-dgx
    environment:
      - MODEL_NAME=BAAI/bge-m3
      # DGX Grace 20-core optimization
      - OMP_NUM_THREADS=20
      - MKL_NUM_THREADS=20
      - PYTORCH_NUM_THREADS=20
      - BLIS_NUM_THREADS=20
      - VECLIB_MAXIMUM_THREADS=20
      - OPENBLAS_NUM_THREADS=20
      # DGX-specific optimizations
      - PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0
      - PYTORCH_ENABLE_MPS_FALLBACK=1
      - CUDA_VISIBLE_DEVICES=""
      - GT2_PLATFORM=dgx
      - MALLOC_ARENA_MAX=8
      - USE_ONNX_RUNTIME=true
      # Grace architecture optimizations
      - CFLAGS=-march=armv8.2-a+fp16+rcpc+dotprod -O3
      - CXXFLAGS=-march=armv8.2-a+fp16+rcpc+dotprod -O3
    deploy:
      resources:
        limits:
          memory: 32G
        reservations:
          memory: 16G
    labels:
      - "gt2.platform=dgx"
      - "gt2.architecture=grace-arm"