0

Apparently tensorflow needs cudnn6 to use the gpu. I have cuda 8 and cudnn5 and 7 but not 6. I get this error whenever I try to import tensorflow after installing tensorflow-gpu: 

ImportError: libcudnn.so.6: cannot open shared object file: No such file or directory

I have cuda-8.0.61-2 installed from the repositories.

So I got cudnn6 from here: https://aur.archlinux.org/packages/cudnn6/

And tried make -si but I get: 

==> Making package: cudnn6 6.0.21-2 (Mon Nov 27 10:06:19 CET 2017)
==> Checking runtime dependencies...
==> Installing missing dependencies...
error: target not found: cuda-8.0
==> ERROR: 'pacman' failed to install missing dependencies.

Any idea how could I fix this?

talonmies
  • 70,661
  • 34
  • 192
  • 269
Atirag
  • 1,660
  • 7
  • 32
  • 60

1 Answers1

1

*. Uninstall cuda from repository, install https://aur.archlinux.org/packages/cuda-8.0/

or

*. Delete dependensi to cuda 8 from cudnn 6 pkg file, "depends=('cuda-8.0')"

I prefer solution 1 becouse it not will break the package dependency

M Hendra Herviawan
  • 133
  • 1
  • 5
  • Thank you. I found a better solution for cudnn 6. On the comments of the AUR package page it says that the PKGBUILD is broken and there's a link for a new one that works correctly – Atirag Dec 01 '17 at 15:13