When trying to build this package for an NVIDIA GTX 1080 Ti, compilation completed, but the kernels failed to execute at run time.

I tracked this down to the following line which causes the kernels only to be built for Compute Capabilities 7.0 and 8.0.

Building for a specific compute capability can be achieved by adding the CUDA_COMPUTE_CAPABILITY definition to the build_slines.sh script. For example, on my system I want compute capability 6.1, so I added -DCUDA_COMPUTE_CAPABILITY=61

# configure
cmake -DCMAKE_INSTALL_PREFIX=${install_dir} \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_C_COMPILER=gcc \
      -DCMAKE_CXX_COMPILER=g++ \
      -DPYTHON_EXECUTABLE=$(which python) \
      -DCUDA_COMPUTE_CAPABILITY=61 \
      ..

Recent CMAKE also seems to have a CMAKE_CUDA_ARCHITECTURES property, but using that one did not seem to work: the CUDA_COMPUTE_CAPABILITY defaults ended up getting used instead.