CuPy supports several "satellite" CUDA libraries: cuTENSOR, NCCL, and cuDNN. Historically CuPy's binary packages have been built with specific versions of these libraries, and we were asking users to install that version of libraries using cupyx.tools.install_library tool, which downloads library archives from NVIDIA website and extract them under ~/.cupy.

However, today these libraries are also available on PyPI (e.g. https://pypi.org/project/cutensor-cu12/ and https://pypi.org/project/nvidia-nccl-cu12/), and it is becoming common to use the PyPI installation among CUDA/Python projects (such as PyTorch). And now that CuPy supports loading both cuTENSOR and NCCL from PyPI wheels (#9201), it does not make much sense to continue providing the tool. Note that cuDNN support is going to be removed as discussed in a separate RFC issue (#8215).

Therefore I am thinking of removing the tool in CuPy v14. I believe this doesn't impact most of the use-cases but will keep this RFC issue open for a while to allow anyone to jump in the discussion.