# High-performance GPU configuration for Animagine MCP
# This compose file includes advanced GPU optimizations
services:
animagine-mcp:
build:
context: .
dockerfile: Dockerfile
# Enable BuildKit for faster, parallel builds
args:
- BUILDKIT_INLINE_CACHE=1
container_name: animagine-mcp-server
runtime: nvidia
# Ports
ports:
- "8000:8000"
# GPU Support (NVIDIA) - Multiple GPU support
deploy:
resources:
reservations:
devices:
# For single GPU (most common)
- driver: nvidia
device_ids: ['0']
capabilities: [gpu]
# For multi-GPU setups, add more entries:
# - driver: nvidia
# device_ids: ['1']
# capabilities: [gpu]
# Volumes
volumes:
# Persist model checkpoints
- ./checkpoints:/app/checkpoints:rw
# Persist LoRA models
- ./loras:/app/loras:rw
# Persist generated outputs
- ./outputs:/app/outputs:rw
# Cache for Hugging Face models (persistent across restarts)
- hf_cache:/root/.cache/huggingface
# Cache for torch models (persistent across restarts)
- torch_cache:/root/.cache/torch
# Environment variables - GPU Optimized
environment:
# Python settings
PYTHONUNBUFFERED: "1"
PYTHONDONTWRITEBYTECODE: "1"
# Hugging Face settings
HF_HOME: /root/.cache/huggingface
TORCH_HOME: /root/.cache/torch
# CUDA Configuration
CUDA_VISIBLE_DEVICES: "0" # Change to "0,1,2" for multi-GPU
CUDA_LAUNCH_BLOCKING: "1"
CUDA_DEVICE_ORDER: "PCI_BUS_ID"
TORCH_CUDA_ARCH_LIST: "7.0,7.5,8.0,8.6,8.9,9.0"
# NVIDIA Driver Configuration
NVIDIA_VISIBLE_DEVICES: "all"
NVIDIA_DRIVER_CAPABILITIES: "compute,utility"
# PyTorch Optimization
TORCH_CUDNN_BENCHMARK: "1" # Enable cuDNN auto-tuner
TORCH_CUDNN_DETERMINISTIC: "0" # Allow non-deterministic for speed
TORCH_CUDA_DEVICE_BATCH_SORT_BACKEND: "nccl"
OMP_NUM_THREADS: "8" # Adjust based on CPU cores
# Memory Optimization
PYTORCH_CUDA_ALLOC_CONF: "max_split_size_mb:512"
# Optional: Disable telemetry
HF_HUB_DISABLE_TELEMETRY: "1"
# Optional: Skip model verification on startup (model is pre-downloaded in image)
# SKIP_MODEL_DOWNLOAD: "true"
# Restart policy
restart: unless-stopped
# Logging
logging:
driver: "json-file"
options:
max-size: "10m"
max-file: "3"
# Enhanced healthcheck - verifies GPU functionality
healthcheck:
test: ["CMD", "python", "-c", "import torch; import sys; print('=== GPU Status ==='); print(f'CUDA Available: {torch.cuda.is_available()}'); print(f'GPU Count: {torch.cuda.device_count()}'); print(f'Current Device: {torch.cuda.current_device()}'); print(f'GPU Name: {torch.cuda.get_device_name()}'); sys.exit(0 if torch.cuda.is_available() else 1)"]
interval: 30s
timeout: 15s
retries: 3
start_period: 60s
volumes:
hf_cache:
driver: local
torch_cache:
driver: local
