FS#61107 - [python-pytorch-cuda] Didn't use the latest cudnn path
Attached to Project:
Community Packages
Opened by laper (laper) - Monday, 17 December 2018, 08:41 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Wednesday, 19 December 2018, 08:55 GMT
Opened by laper (laper) - Monday, 17 December 2018, 08:41 GMT
Last edited by Sven-Hendrik Haase (Svenstaro) - Wednesday, 19 December 2018, 08:55 GMT
|
Details
Description:
In PKGBUILD, cudnn path was exported as: export CUDNN_LIB_DIR=/opt/cuda/lib64 export CUDNN_INCLUDE_DIR=/opt/cuda/include But in cudnn 7.4.1.5-2, files were linked to /usr/include/cudnn.h and /usr/lib/libcudnn.so. Without building with cudnn, model training process would much slower. Solution: export CUDNN_LIB_DIR=/usr/lib export CUDNN_INCLUDE_DIR=/usr/include |
This task depends upon
Closed by Sven-Hendrik Haase (Svenstaro)
Wednesday, 19 December 2018, 08:55 GMT
Reason for closing: Fixed
Wednesday, 19 December 2018, 08:55 GMT
Reason for closing: Fixed
It's python-pytorch-cuda 1.0.0-2.
Sorry for such bug report, this is my first time bug report. Came to this bug page through the package actions, thought this would automatically related to the package, like github issue, but it's not :(
Anyway, the package got a pkgrel bump for the new cudnn, but no changes to those variables which still point to /opt, so, that does seem to be a problem: https://git.archlinux.org/svntogit/community.git/commit/trunk?h=packages/python-pytorch&id=4463b688805d6bb0360d83a268912002f3ef16f4