cupy currently requires development headers that aren't in the "runtime" CUDA Docker images (https://github.com/cupy/cupy/issues/8466). It seems like eventually the plan is to support CUDA from pip (https://github.com/cupy/cupy/issues/8013), and indeed the headers are already available from the nvidia-cuda-runtime-cu12 on PyPI, but let's not mix and match CUDA sources unless we need to.
(Also, let's use Rocky instead of Ubuntu, since most users will probably be on a GPU cluster running a RHEL variant.)
cupy currently requires development headers that aren't in the "runtime" CUDA Docker images (https://github.com/cupy/cupy/issues/8466). It seems like eventually the plan is to support CUDA from pip (https://github.com/cupy/cupy/issues/8013), and indeed the headers are already available from the
nvidia-cuda-runtime-cu12
on PyPI, but let's not mix and match CUDA sources unless we need to.(Also, let's use Rocky instead of Ubuntu, since most users will probably be on a GPU cluster running a RHEL variant.)