From 79a77880e5d07ece235e8a0fa0a46a1c204ec5d7 Mon Sep 17 00:00:00 2001 From: Jinuk Moon Date: Sun, 19 Jul 2026 16:32:57 +0900 Subject: [PATCH] docs: full CUDA toolkit (NVRTC dev libs) required to build the extension A compiler-only toolchain (e.g. conda's cuda-nvcc alone) fails the CMake build with 'CUDA::nvrtc but the target was not found' (CMakeLists.txt links CUDA::nvrtc). Document the requirement and the CUDAToolkit_ROOT workaround, as suggested in #207. --- docs/installation.rst | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-) diff --git a/docs/installation.rst b/docs/installation.rst index 5ade5c0c..19610d55 100644 --- a/docs/installation.rst +++ b/docs/installation.rst @@ -36,7 +36,7 @@ You need the following to install OpenEquivariance: c++ --version - To build the extension with an alternate compiler, set the + To build the extension with an alternate compiler, set the ``CC`` and ``CXX`` environment variable and retry the import: @@ -46,9 +46,23 @@ You need the following to install OpenEquivariance: export CXX=/path/to/your/g++ python -c "import openequivariance" + Building the extension also requires a **full CUDA toolkit**, including + the NVRTC development libraries (the CMake build links ``CUDA::nvrtc``). + A compiler-only toolchain — for example conda's ``cuda-nvcc`` package on + its own — fails with:: - These configuration steps are required only ONCE after - installation (or upgrade) with pip. + target_link_libraries ... CUDA::nvrtc but the target was not found + + In that case, point CMake at a full toolkit installation and retry the + step that failed (the ``pip install`` for a source build, or the import): + + .. code-block:: bash + + export CUDAToolkit_ROOT=/usr/local/cuda + python -c "import openequivariance" + + These configuration steps are required only ONCE after + installation (or upgrade) with pip. .. tab:: JAX NVIDIA GPUs