Cuda 9.1. not compatible with nvidia 384

drivers
nvidia

#10

It’s not that big of a deal, not everyone uses the same setup and things get missed.

Just use downgrade to install cuda 9.0 and don’t upgrade to cuda 9.1 while Manjaro uses the nvidia 384 branch.

If you don’t have cuda 9.0 you can download it from the ALA and move it to your pacman cache folder, then use downgrade to install it.

I’m not going to explain how to enable ALA downloading via downgrade for stable branch and you would do this at your own risk.

https://archive.archlinux.org/packages/c/cuda/


#11

I didn’t know. I thought that the team had more control over the packages. Of course in such case I will take care.


#12

Measures taken, everything works fine. Is there a way to disable the update of a certain package? Because Manjaro package manager keeps signalling red.


#13

We have full control over the packages in the repos, but you can’t “hold” a package that has already been released. Once they reach the stable branch and have been installed by users it’s a bit late to start wanting to hold them back.

We could withdraw them, but that causes its own problems.


#14

Once you run downgrade it’ll ask if you want to ignore updates for cuda, just select yes and you’ll be all set.


#15
/etc/pacman.conf

Edit the IgnorePkg line and add whatever packages you want pacman to not update.

IgnorePkg = cuda

#16

Yeah, that I did. Thank you.


#17

Also, be sure to install a version of cuDNN which is compatible with the version of CUDA your are downgrading to.

In my case, I reinstalled cuda-9.0.176-4 from https://archive.archlinux.org/packages/c/cuda/ and cudnn-7.0.4-2 from https://archive.archlinux.org/packages/c/cudnn/, then I installed tensorflow and tensorflow-gpu using pip and everything started working fine again.

I am on Manjaro 17.1.3 with a GTX 1070 and NVIDIA drivers version 384.111.


#18

Where can we submit unit tests to automate the testing of this dependency?


#19

How would unit tests help?


#20

A simple continuous integration profile with the following test would have easily caught this issue:

  1. Add new CUDA version to manjaro
  2. Compile a simple CUDA helloworld application via Jenkins / Travis etc
  3. Run code, check return code.

It should not be up to the user (or developer ) to validate such a trivial dependency decoupling.


#21

That’s not a unit test, that’s CI.

If you’d like, set up a PoC to demonstrate how it would work and how it would scale to all packages in the repos.


#22

It appears the 390 drivers are in stable now, so removing “IgnorePkg = cuda” should be safe?


#23

Only one way to find out …


#24

I have been using it in testing since day one with no issue :wink:


#25

I’m scared to change to testing. I’m not that experienced with Manjaro/Arch/Linux. :slight_smile:


#26

If you’re not running Gnome as your desktop environment, then it should be save. We are still testing stuff, to make it smooth as possible.


#27

Thanks. :slight_smile: I actually don’t even use my machine via desktop more than a few times a year, it’s a server I run 24/7 for some programming projects and to operate Go playing AI bots on a server. I’ll go ahead and update everything this morning, doesn’t sound like I have anything to worry about.


#28

It is always good to read the update announcements first. We always give tips and tell what might not run properly. Please look at these at least:


#29

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.