docker-compose.gpu.yml

# docker-compose.gpu.yml
# Optional overlay for GPU acceleration (NVIDIA)
#
# This overlay enables GPU runtime for celery-worker on systems with NVIDIA GPUs.
# It is automatically detected and loaded by opentr.sh when NVIDIA Container Toolkit
# is available.
#
# Usage:
#   ./opentr.sh start dev              # Auto-detects GPU and applies if available
#   OR manually:
#   docker compose -f docker-compose.yml -f docker-compose.override.yml -f docker-compose.gpu.yml up
#
# Note: This file is automatically included by opentr.sh when NVIDIA GPU is detected.
# On macOS or systems without NVIDIA GPU, this file is not loaded, allowing the base
# docker-compose.yml to work on CPU-only systems.
#
# Configuration:
#   GPU_DEVICE_ID       - GPU device for AI work in normal mode (default: 0)
#   GPU_SCALE_DEVICE_ID - GPU device for AI work in --gpu-scale mode (default: 2)
#   Set in .env file: GPU_DEVICE_ID=1

services:
  celery-worker:
    # Enable NVIDIA runtime for GPU acceleration
    # Note: deploy.resources.reservations is preferred over runtime: nvidia
    # in newer Docker versions (20.10+) for better compatibility
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              device_ids: ['${GPU_DEVICE_ID:-0}']

  celery-cpu-worker:
    # GPU access for nvidia-smi stats collection and CUDA-accelerated clustering.
    # Uses 'count: all' so it can query ANY GPU device — this is needed because:
    # 1. In --gpu-scale mode the active AI GPU is GPU_SCALE_DEVICE_ID (e.g. 2)
    #    which differs from GPU_DEVICE_ID (e.g. 0)
    # 2. Speaker clustering uses GPU_CLUSTERING_DEVICE (falls back to GPU_DEVICE_ID)
    #    for PyTorch fp16 similarity matrix computation + AHC clustering
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              count: all