Run:
uv run python scripts/check_environment.pyIf uv is missing, install uv and rerun setup.
GET /models will show stable_audio_python.available=false when stable_audio_3
cannot be imported. Run:
./scripts/install_python_provider.shMedium may need CUDA and Flash Attention. Small Music and Small SFX are the safer first local tests.
If loading a Python model returns 401 Client Error for a stabilityai/stable-audio-3-*
Hugging Face URL, the package is installed but the model repo is gated. Accept the
model terms on Hugging Face and log in with a read token:
uv run hf auth loginThen check from the server:
curl "http://127.0.0.1:5178/huggingface/status?check_models=true"If the model status is requires_approval_or_login, open the corresponding
Hugging Face model page, accept the terms, and rerun the check.
On Apple Silicon, use stable_audio_mlx first; it is the verified local route for
this project.
The MLX provider requires Apple Silicon and the official optimized/mlx/sa3 CLI.
./scripts/install_mlx_provider.shIf the official repo is outside this project:
export GERM_MLX_REPO_DIR=/path/to/stable-audio-3Generation failures still create metadata under output/metadata. Open the JSON and
check error, request, stdout, and stderr.
Start the server first:
./scripts/run_server.shThen launch the dashboard:
./scripts/run_dashboard.shConfirm the dashboard Server URL is:
http://127.0.0.1:5178
Review the relevant Stability AI model license, API terms, and any Hugging Face access terms before commercial use or redistribution. This project keeps providers replaceable so local, API, or enterprise backends can be selected according to the deployment requirements.