1

I'm using cupy with Spyder3.3.6 and Python 3.7.5 in Windows machine(Win10 Pro 64bit, i7-7700, 8GBMemory, GTX-1060-6GB).
Version of cupy, chainer, cuda and cuDNN are 6.0.0, 5.3.0, 10.1.243, and 7.6.4,respectively.

When I import cupy, this error has occured:

C:\Users\(username)\Anaconda3\envs\gpgpu\lib\site-packages\chainer\backends\cuda.py:98: UserWarning: cuDNN is not enabled.
Please reinstall CuPy after you install cudnn
(see https://docs-cupy.chainer.org/en/stable/install.html#install-cudnn).
  'cuDNN is not enabled.\n'

However, try

import cupy

again, the error hasn't occur and cupy works well. (i.e. write "import cupy" 2 time in the code)
However cupy is actually working, it fells strange to me.
I already tried to re-install related packages, but the problem hasn't solved.

How I can solve the problem?

mas
  • 35
  • 7

1 Answers1

1

First of all, it seems you are using different versions of chainer/cupy. We recommend chainer and cupy to match the version number as we develop them in tandem.

How did you install CuPy? We support pre-built wheels for windows, which include cudnn and nccl versions that we guarantee to work. You can install them with pip install cupy-cuda101.

emcastillo
  • 306
  • 1
  • 4
  • Thanks for your answer.
    I installed CuPy from conda. Can I install pre-vuilt wheels for windows via conda instead of pip?     – mas Dec 02 '19 at 10:08
  • 1
    Conda builds do not include cudnn binaries and they resort to the conda dependence. Also, chainer versions in conda are outdated. At this moment we recommend installing chainer/cupy using pip if possible. You can always try to build cupy and chainer from source if pip is not an option. – emcastillo Dec 03 '19 at 03:10