Do not bundle cuDNN v8 for CUDA 10.1/10.2/11.0 packages

We’re now discussing removal of bundled cuDNN in CuPy wheels hosted on PyPI. This is because cuDNN v8 is quite huge (1.5GB) and not reasonable to redistribute on our own.

This affects:

• CUDA 10.1/10.2 wheels for CuPy v8
• CUDA 11.0 wheels for CuPy v7 and v8

Requirements

1. Load shared libraries placed outside of the package
2. (optionally) Provide means to download cuDNN automatically

Implementation

Requirement 1

• Before loading cupy_backends,
  • Find the library from search paths in the following order:
    • ~cupy/_data/lib (shared libaries bundled in wheels)~ update: this is not needed as cuDNN binary is not located here
    • CUPY_LIBRARY_PATH env var
    • default system search path
  • If the library is found, dlopen the library. Otherwise, show a warning (please install the lib manually) to users on the initial use of each library.

Requirement 2

• Users run python -m cupyx.utils.download_cudnn cuda_version
• Download cuDNN binary to e.g. ~/.cupy/cuda_libs/{cuda_version}/{library_name}/{library_version}/lib/libcudnn.so.8.0.2

If we provide this implementation, we should add ~/.cupy/cuda_libs/{cuda_version}/{library_name}/{library_version}/lib to search path in requirement 1.

[kmaehashi]

can we add https://github.com/unnonouno/cudnnenv as a requirement and call it from setup.py?

I come up with that idea too, but cudnnenv does not support Windows. Also, wheel installations does not invoke setup.py, so cudnn download should be done by users (or maybe automatically by CuPy)

[leofang]

@jakirkham, I don’t think Anaconda has packaged cuDNN v8, nor would this have impact to conda releases (conda’s cuDNN would be found when searching “default system path” in Requirement 1). Just FYI.