# sd-server configuration -- EnvironmentFile for sd-server.service
# After changes:  systemctl restart sd-server

# Base directory for model files.  Configurable; on some deployments this may
# point to a path under /home or a separate mount (requires a systemd drop-in
# for ReadOnlyPaths/ProtectHome, see the .service file).
SD_MODELS_DIR=/var/lib/sd-cpp/models

# Path to the diffusion model to load (MUST be set before starting the service).
# Supported formats: .gguf (quantized), .safetensors, .pt, .pth
# For GGUF quantized models use the .gguf path directly.
# For original weights sd.cpp loads .safetensors/.pt natively.
SD_MODEL=${SD_MODELS_DIR}/model.safetensors

# Bind address -- keep internal, control access via nftables or reverse proxy.
# Upstream default port is 1234; we use 8080 here for consistency with other
# inference servers.  Adjust to taste.
SD_HOST=127.0.0.1
SD_PORT=8080

# CPU threads -- conservative default of 2.  MUST be adjusted to the machine!
# Rule of thumb: number of physical cores (NOT SMT/hyperthreads).
# sd.cpp's GGML backend scales with physical cores, not with hyperthreads.
SD_THREADS=2

# Extra arguments passed to sd-server, for example:
#   --diffusion-model /path/to/unet.safetensors
#   --vae /path/to/vae.safetensors
#   --lora-model-dir /var/lib/sd-cpp/loras
#   --verbose
#   --flash-attn
SD_EXTRA_ARGS=