FS#60353 - [python-pytorch-cuda] 1.0rc1-2 compiled without cuDNN support
Attached to Project:
Community Packages
Opened by Deepak Subburam (subburam) - Monday, 08 October 2018, 19:14 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Wednesday, 17 October 2018, 08:09 GMT
Opened by Deepak Subburam (subburam) - Monday, 08 October 2018, 19:14 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Wednesday, 17 October 2018, 08:09 GMT
|
Details
Description:
cuDNN libraries not used by pytorch, and triggers the following warning, running about 50% slower than prior version: "/usr/lib/python3.7/site-packages/torch/backends/cudnn/__init__.py:89: UserWarning: PyTorch was compiled without cuDNN support. To use cuDNN, rebuild PyTorch making sure the library is visible to the build system." Steps to reproduce: In a python shell, do import pytorch.nn rnn = torch.nn.RNN(100,100).cuda() |
This task depends upon
Closed by Sven-Hendrik Haase (Svenstaro)
Wednesday, 17 October 2018, 08:09 GMT
Reason for closing: Fixed
Wednesday, 17 October 2018, 08:09 GMT
Reason for closing: Fixed
Comment by Cat (lasercat) -
Tuesday, 16 October 2018, 08:39 GMT
Comment by
Sven-Hendrik Haase (Svenstaro) -
Tuesday, 16 October 2018, 21:26 GMT
Comment by
Deepak Subburam (subburam) -
Wednesday, 17 October 2018, 00:12 GMT
Adding export WITH_CUDNN=1 after line 128 to the PKGBUILD fixed
this on our side.
Ah yeah that must have happened when I cleaned up the PKGBUILD.
Please test the new package.
Yes, warning no longer occurs with the new package. So, fixed.
Thanks!