Installation

RayD is published on PyPI as prebuilt CUDA wheels. You do not need the CUDA toolkit or the OptiX SDK to install it; both are only needed to build from source.

Requirements#

The wheels contain native code for sm_70, sm_75, sm_80, sm_86, sm_87, sm_89, sm_90, sm_100, sm_101, and sm_120, plus compute_120 PTX for newer GPUs. They are built with CUDA 12.8 and the OptiX 8.1 API.

OptiX itself ships inside the NVIDIA driver. If the driver cannot create an OptiX context, a scene created with the default trace_backend="auto" uses RayD's pure-CUDA BVH instead, so every query still runs. macOS and CPU-only machines are not supported.

Install#

There are three distributions, all at version 0.8.0:

pip install rayd-drjit==0.8.0

Verify the installation#

Run this for the Dr.Jit backend:

python -c "import rayd.drjit as rd; print(rd.optix_available(), rd.device_count()); print(rd.Scene().capabilities()['trace_backend'])"

Expected output on a machine with a working OptiX driver:

True 1
optix

cuda in the last line means OptiX is not available and RayD uses its CUDA fallback. Dr.Jit may print jitc_llvm_init(): LLVM API initialization failed on import. That message is about Dr.Jit's CPU backend, which RayD does not use, and you can ignore it.

Run this for the PyTorch backend:

python -c "import rayd.torch as rt; print(rt.backend_capabilities()['backend'], rt.backend_capabilities()['torch_compile'])"

torch True

Then run the quickstart, which checks a real intersection and its gradient.

Environment variables#

Build from source#

You need a source build only to change RayD itself or to link it into a native CMake project. The requirements are Python 3.10 to 3.14, CMake 3.22 or newer (3.26 for the Torch backend), a C++17 compiler (Visual Studio 2022 with Desktop C++ tools on Windows), the CUDA toolkit, and the OptiX SDK headers.

git clone https://github.com/Asixa/RayD.git
cd RayD
python -m pip install cmake ninja scikit-build-core
python -m pip install "drjit==1.3.1" "nanobind==2.11.0"
python -m pip install --no-build-isolation -ve drjit

For the Torch backend, install torch==2.10.0 from the cu128 index and run python -m pip install --no-build-isolation -ve torch. On Windows, scripts\build_local.cmd -Backend drjit (or -Backend torch) builds only the architecture of the GPU it detects, which is much faster than the release matrix.

Troubleshooting#