Installation

make sure the CUDA_HOME environment variable points to the cuda toolkit directory (e.g /usr/local/cuda).

for linux

pip install .

or

pip install . --no-build-isolation --no-deps

for GH200 on top of existing container/uenv (cuda release wheels don't exist for aarch64, so need pip to build in isolation).