Do not bundle cuDNN / NCCL for all wheel packages starting in CuPy v9

We have been distributing binary packages (cupy-cudaXXX) via PyPI. However recently the size of GPU-related packages (including CuPy) is starting to cause a problem on PyPI (see discussions on Python forum).

We take this problem seriously, and to help PyPI ecosystem healthy, we are planning to stop bundling cuDNN / NCCL shared libraries from all wheels, starting in v9 releases. Note that cuDNN / NCCL are still supported and enabled in wheels, but users who want to use these features need to install the library via python -m cupyx.tools.install_library command.

---

FYI, here are our past efforts to reduce the package size on PyPI:

• CuPy v7: Do not bundle cuDNN v8.x (#3724)
• CuPy v8: Remove outdated pre-release wheels (#4360)

[pentschev]

As we discussed earlier today on our call, could we extend this effort to conda packages, effectively making cuDNN an optional install rather than a requirement?

[kmaehashi]

Currently, there’s no way to programmatically force building without cuDNN when it is available.

You can comment out this line: https://github.com/cupy/cupy/blob/704f87e927998f31534ab76b5d831f967114a84c/install/cupy_builder/_modules.py#L381

or, you can make a faulty dummy header so that CuPy does not recognize cuDNN as available.

echo "#error dummy" > cudnn.h
CFLAGS="-I${PWD}" pip install -v .