Installation

pip install "kanforge @ git+https://github.com/fx-xf/kanforge"          # NumPy reference only
pip install "kanforge[torch] @ git+https://github.com/fx-xf/kanforge"   # + PyTorch layers (CUDA, ROCm, MPS, CPU)

Once the package is on PyPI, pip install "kanforge[torch]" will do the same.

Optional extras, combine as needed (pip install "kanforge[torch,export] @ git+https://github.com/fx-xf/kanforge"):

extra	adds
torch	PyTorch layers, kernels (Metal on MPS; Triton on CUDA needs triton)
mlx	MLX layers and Metal kernels (Apple silicon)
jax	JAX / Flax NNX layers and Pallas kernels (Python ≥ 3.11)
symbolic	SymPy, for formulas and the kanpiler
plot	matplotlib, for edge plots
hf	transformers, for kanify
export	onnx, onnxruntime
coreml	coremltools (macOS)
test	everything the test suite uses

Requirements: Python ≥ 3.10, NumPy ≥ 1.26. Kernels: PyTorch ≥ 2.6 for MPS (torch.mps.compile_shader), Triton ≥ 3.0 for CUDA.