Helion

Helion is a small LLM training stack built around two layers:

• tritium: handwritten Triton kernels for training and inference primitives.
• helion: a compact PyTorch-like layer over those kernels for modules, optimizers, and training utilities.

Project priorities, in order:

1. Cleanliness
2. Optimization
3. Being featureful

The codebase intentionally favors a small, readable API surface over broad coverage. New features should start with clear interfaces and tests before additional tuning.

Layout

• tritium/ops/: Triton-backed primitive operations.
• tritium/: public kernel package exports and compatibility shims.
• helion/: module, optimizer, and training abstractions built on Tritium.
• tests/: CUDA correctness tests for Tritium primitives and Helion wrappers.
• examples/: small training scripts and reference comparisons.

Development

Install the package in editable mode with development dependencies:

pip install -e ".[dev]"

Run the standard checks:

ruff check .
ruff format --check .
pytest -q

Most tests require CUDA and are skipped automatically when CUDA is unavailable.