Installation
greyhound requires Python >= 3.11, PyTorch >= 2.7.1, and a CUDA-capable GPU.
Installing with pip
Section titled “Installing with pip”greyhound-kernels is available on PyPI.
It provides the greyhound Python import package:
pip install greyhound-kernelsOr with uv:
uv add greyhound-kernelsInstalling from source
Section titled “Installing from source”Clone the repository:
git clone https://github.com/tyler-romero/greyhound.gitcd greyhoundThen install in editable mode with development dependencies:
uv sync --all-extraspip install -e ".[dev]"Optional dependencies
Section titled “Optional dependencies”For the Quack dependency used by bonus Newton-Schulz utilities and Quack comparison benchmarks:
uv sync --extra quackpip install -e ".[quack]"For the full benchmark provider set:
uv sync --extra dev --group thirdpartyThis uv-only dependency group installs comparison providers used by benchmark scripts, such as
Dion,
liger-kernel,
gram-newton-schulz,
flash-linear-attention, and
causal-conv1d. The group is intentionally not published as a Greyhound package
extra because PyPI rejects direct Git dependencies in uploaded package metadata.
For Modal benchmark execution:
uv sync --extra modalpip install -e ".[modal]"