Development

mise provisions the toolchain (mise.toml: Python 3.12 + uv) and runs every project command (mise run <task>). The runtime targets the Linux/CUDA environment documented in the inference guide; the lock is restricted to sys_platform == 'linux' via required-environments. The quality and documentation groups can be installed without model dependencies on a Linux host.

mise install
mise run install
mise run check
mise run test
mise run build

mise run install installs the runtime, training, agent and telemetry extras, and quality tools. The site lives in site/ (Nuxt + UnoCSS): mise run site-preview. Run mise tasks to see the available commands.

Dependencies

pyproject.toml owns the package metadata, dependencies, tool settings, and uv configuration. uv.lock records the resolution for Python 3.12 on Linux, including the pinned Bub Git revision and the +cu128 PyTorch artifacts bound through tool.uv.sources and the explicit pytorch-cu128 index. The published experiment's environment snapshot remains historical evidence, not an installation input.

After changing dependencies, run mise run lock, review the lock diff, and sync the relevant selections with uv sync. CI refuses a missing or stale lock rather than resolving different versions. Avoid replacing a trained run's environment while it is active.

Checks and builds

mise run check verifies the lock, repository hygiene, Ruff lint and formatting, and ty checks for src against the installed dependencies. mise run lint runs the checks that only need the quality group. uv run --group quality prek install optionally installs the same hygiene hooks locally; installing hooks is not required to run the checks.

mise run test runs pytest, including the pinned RLCD loss and gradient values, prefix-cache updates and gradients, calibration metrics, and data-split isolation. It does not load model weights, download data, or prove GPU training/inference acceptance. CI runs these tests with Hugging Face access disabled. GPU acceptance remains a separate step for changes to execution or training.

Run an individual file or case with uv run pytest tests/test_rlcd.py or uv run pytest -k gradients. Test discovery is limited to tests/ in pyproject.toml.

The test job installs the same locked CUDA wheels, checks types, and runs the tests on CPU. Quality and website jobs do not install the model runtime.

mise run build uses uv_build to produce a wheel and source distribution in dist/. Local model weights, datasets, caches, and experiment runs are excluded. The CI build does not publish packages. Website validation and deployment use the separate website workflow.

The mise tasks and separated quality/test jobs adapt the relevant parts of Bub's CI layout and CI.