Thank you for contributing! This repo uses:
- Typer CLI as the primary interface
- Ruff for lint/format
- Pytest for tests (fast tests only in CI)
pip install -e '.[dev]'
pre-commit install
- Run fast tests:
pytest -q tests -k 'not slow'
- Lint/format:
ruff check . && ruff format --check .
README/plans/ is gitignored by default (local implementation plans). To commit an updated plan, use git add -f README/plans/<file>.md.
Caption/VQA training on a real gated multimodal checkpoint is covered by tests/test_smoke_image_multimodal.py (requires Hub auth). Locally:
HF_TOKEN=... pytest tests/test_smoke_image_multimodal.py -m integration -vRepository maintainers can add a HF_TOKEN (read) secret and run the Gated multimodal image smoke workflow manually (workflow_dispatch). If the secret is missing, those tests skip and the workflow still passes.
- Include a short description and screenshots/logs for UX changes
- Keep edits focused; avoid unrelated formatting churn
- Green CI required