Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
80 lines (56 loc) · 1.9 KB

CONTRIBUTING.md

File metadata and controls

80 lines (56 loc) · 1.9 KB

Contributing to TIDE

Thanks for your interest in TIDE! Here's how to contribute.

Development Setup

git clone https://github.com/RightNowAI/TIDE.git
cd TIDE
pip install -e ".[test]"    # installs with CUDA kernels if GPU available

No GPU locally? Use TIDE_NO_CUDA=1 pip install -e ".[test]" for CPU-only.

Running Tests

# CPU-only tests (fast, no GPU)
pytest tests/ -k "not cuda and not kernels" -v

# Full suite (requires CUDA GPU)
pytest tests/ -v

# Cloud GPU tests
modal run modal_setup/ci_app.py

Code Style

We use ruff for linting:

pip install ruff
ruff check python/ tests/

Project Layout

  • python/TIDE/ -- Python package (runtime, calibration, adapters)
  • csrc/ -- CUDA kernels (C++/CUDA)
  • tests/ -- Test suite
  • examples/ -- Example scripts
  • modal_setup/ -- Cloud GPU infrastructure
  • benchmarks/ -- Performance benchmarks

Adding Support for a New Model Architecture

Most models work automatically via UniversalAdapter. If your model doesn't, add a built-in adapter:

  1. Create python/TIDE/adapters/mymodel.py implementing BaseAdapter
  2. Register it in python/TIDE/adapters/auto.py under ADAPTER_REGISTRY
  3. Add a test in tests/test_adapters.py

CUDA Kernel Development

Kernels are in csrc/kernels/. All kernels:

  • Are templated on scalar_t (float, __half, __nv_bfloat16)
  • Use dtype_utils.cuh for load/store helpers
  • Expose named C entry points (not templates) for linking
  • Accumulate in float32 for numerical stability

After modifying kernels, rebuild: pip install -e .

Pull Requests

  1. Fork and create a branch
  2. Make your changes
  3. Run pytest tests/ -v (at least CPU tests)
  4. Run ruff check python/ tests/
  5. Open a PR with a clear description

Reporting Issues

Please include:

  • Model name and size
  • GPU type
  • PyTorch and transformers versions
  • Full error traceback