services:

heartmula:

# Use pre-built image from GitHub Container Registry (faster)

image: ghcr.io/fspecii/heartmula-studio:latest

# Or build locally: uncomment the lines below and comment out the image line above

# build:

# context: .

# dockerfile: Dockerfile

container_name: heartmula-studio

# GPU access - requires NVIDIA Container Toolkit

deploy:

resources:

reservations:

devices:

- driver: nvidia

count: all # Use all available GPUs

capabilities: [gpu]

ports:

- "8000:8000"

# Allow container to reach host services (Ollama)

extra_hosts:

- "host.docker.internal:host-gateway"

volumes:

# Persist models (auto-downloaded on first run, ~5GB)

- ./backend/models:/app/backend/models

# Persist generated music (accessible from host)

- ./backend/generated_audio:/app/backend/generated_audio

# Persist reference audio uploads

- ./backend/ref_audio:/app/backend/ref_audio

# Persist database (song history) - uses named volume

- heartmula-db:/app/backend/db

environment:

# GPU auto-detection (recommended)

- HEARTMULA_4BIT=auto

- HEARTMULA_SEQUENTIAL_OFFLOAD=auto

# PyTorch memory optimization

- PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

# HuggingFace cache location

- HF_HOME=/app/backend/models

# Database location (for Docker)

- HEARTMULA_DB_PATH=/app/backend/db/jobs.db

# Ollama on host machine (for AI lyrics generation)

- OLLAMA_HOST=http://host.docker.internal:11434

restart: unless-stopped

healthcheck:

test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]

interval: 30s

timeout: 10s

retries: 3

start_period: 120s # Model loading takes time

# Named volume for database persistence

volumes:

heartmula-db: