services: hf-image-api: build: context: . dockerfile: Dockerfile target: final image: hf-image-api:latest container_name: hf-image-api ports: - "8000:7860" # ── Persistent volumes ───────────────────────────────────────────────────── volumes: # Model weights persist between container restarts (avoids re-downloading) - model_cache:/app/model_cache # Generated images persist and are accessible on the host - ./outputs:/app/outputs # ── Environment (edit here or in .env) ──────────────────────────────────── environment: # Get your free token at https://huggingface.co/settings/tokens # Enables higher rate limits and faster file resolution from HuggingFace HF_TOKEN: ${HF_TOKEN:-} MODEL_ID: ${MODEL_ID:-nota-ai/bk-sdm-small} AUTO_LOAD_MODEL: ${AUTO_LOAD_MODEL:-true} CPU_THREADS: ${CPU_THREADS:-0} ENABLE_ATTENTION_SLICING: ${ENABLE_ATTENTION_SLICING:-true} DEFAULT_STEPS: ${DEFAULT_STEPS:-20} MAX_STEPS: ${MAX_STEPS:-50} DEFAULT_WIDTH: ${DEFAULT_WIDTH:-512} DEFAULT_HEIGHT: ${DEFAULT_HEIGHT:-512} MAX_WIDTH: ${MAX_WIDTH:-768} MAX_HEIGHT: ${MAX_HEIGHT:-768} DEFAULT_GUIDANCE_SCALE: ${DEFAULT_GUIDANCE_SCALE:-7.5} OUTPUTS_DIR: /app/outputs MODEL_CACHE_DIR: /app/model_cache # ── CPU resource limits ──────────────────────────────────────────────────── # Remove or adjust based on your machine deploy: resources: limits: cpus: "4.0" # cap at 4 cores (set to your core count) memory: 8G # 8 GB for bk-sdm-small; use 12 GB for SD v1.5 reservations: memory: 2G restart: unless-stopped healthcheck: test: ["CMD", "curl", "-f", "http://localhost:7860/health"] interval: 30s timeout: 10s retries: 3 start_period: 90s # give extra time for model download on first run # ── Named volumes ────────────────────────────────────────────────────────────── volumes: model_cache: driver: local