os-model-library-setup

Provide a spec file describing the model's library name, package directory, repo URL, dependencies, and GPU requirements. The skill renames the template directory, adds the model repo as a pinned git submodule, updates configuration files, and generates the manifest and advanced library file.

Run git submodule add with the model repo URL inside the package directory, then pin it to a release tag or the current HEAD commit. The pinned SHA guarantees reproducible installs, and the advanced library reinstalls the package when the pinned commit changes.

For submodules without setup.py or pyproject.toml, the advanced library uses a sys.path insertion variant instead of pip install. The commit sentinel becomes the sole install signal, and sys.path is re-applied on every library load since it does not persist across engine restarts.

No. The engine is the host that loads the library, so listing it as a runtime dependency installs a second engine that can shadow the running one. Keep it in the dev dependency group and declare the required version via engine_version in the manifest JSON.

Packages like auto_gptq and flash-attn need torch at build time, but pip's build isolation hides the venv's packages. The generated library installs torch first via the manifest's pip_dependencies and uses --no-build-isolation when installing the submodule's requirements.txt.