Installation

Install Core with pip:

pip install witwin

witwin 0.4.0 is a pure-Python wheel (witwin-0.4.0-py3-none-any.whl) on PyPI. It contains no C++ or CUDA code, so it installs the same way on Windows, Linux and macOS. You usually do not install it by hand: witwin-maxwell, witwin-radar and witwin-channel pull in the version they need.

Requirements#

Core has no Dr.Jit, Mitsuba or native dependency. Every continuous value is a PyTorch tensor, and Core computes on the device where your tensors live.

Choose a PyTorch build#

pip installs the default torch wheel from PyPI if none is present. On Windows that wheel is CPU-only. If you want CUDA tensors, install a CUDA build of PyTorch first, then Core:

python -m pip install torch --index-url https://download.pytorch.org/whl/cu128
python -m pip install witwin==0.4.0

If you plan to use Channel, match its pins instead: CPython 3.11 and torch==2.10.0. See Channel installation.

Verify the installation#

import torch
import witwin.core as core
 
print(core.__version__, torch.__version__)
box = core.Box(size=(2.0, 2.0, 2.0))
print(box.signed_distance(torch.tensor(0.0), torch.tensor(0.0), torch.tensor(0.0)))

Expected output (your PyTorch version string may differ):

0.4.0 2.10.0
tensor(-1.)

The point at the origin is 1 m inside a 2 m cube, so its signed distance is -1.

Install from source#

Clone the repository and install it in editable mode when you work on Core itself:

git clone https://github.com/witwin-ai/witwin-core.git
cd witwin-core
python -m pip install -e .
python -m pytest

The build backend is Hatchling. The source tree contains only Python files.