You are not logged in.

#1 2017-10-28 01:59:20

simon_sjw
Member
From: Sydney, Australia
Registered: 2015-03-20
Posts: 22

CUDA v8 cuDNN v6 dependency issue

Hi All,
I've installed CUDA v8 from the AUR for use with Tensorflow.
This version is one down from the latest release (CUDA v9).
To use Tensorflow, I also need cuDNN v6. (also one down from the latest release cuDNN v7)

There is a version 6 cuDNN in the AUR (thanks petronny) but it is built to be dependent on CUDA v9, the latest release, and not CUDA v8, the version I have to use with Tensorflow.

Can anyone advise? is it possible to override the versioning on install? Would that be the preferred method or should I have a go at repacking cuDNN v6 separately this time dependent on CUDA v8?

(I know cuDNN v6 works fine with both CUDA v8 and CUDA v9).

Simon

Offline

#2 2017-10-28 02:06:48

circleface
Member
Registered: 2012-05-26
Posts: 639

Re: CUDA v8 cuDNN v6 dependency issue

As far as I can tell, cuDNN depends on cuda, not any versioned cuda.  If it isn't working for you, you might want to leave a comment on the aur page asking the maintainer if there are options.

Offline

#3 2017-10-28 11:47:15

Lone_Wolf
Member
From: Netherlands, Europe
Registered: 2005-10-04
Posts: 11,868

Re: CUDA v8 cuDNN v6 dependency issue

[Community] repo has a tensorflow-cuda package that looks like it should work with latest cuda and cudnn , have you tried that ?


Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.


(A works at time B)  && (time C > time B ) ≠  (A works at time C)

Offline

#4 2017-10-29 11:35:23

simon_sjw
Member
From: Sydney, Australia
Registered: 2015-03-20
Posts: 22

Re: CUDA v8 cuDNN v6 dependency issue

Hey circleface/ Lone_Wolf,
Thanks for getting back.
circleface: agreed, I've left a message with them and we'll see what they think. Honestly, it actually boils down to copying three files from the cuDNN package to the right CUDA directories. I can do it manually but wanted to get it right.

Lone_Wolf: the issue is that I'm installing from within Anaconda - Tensorflow would have to be inside the right environment and CUDA is outside that in /opt.

I'll update when I hear back.

Thanks again.

Simon

Offline

Board footer

Powered by FluxBB