2

I've spent two days already trying to get it done, but no joy. Please help!

I'm trying to install CUDA 5.5 drivers, the ones that come with the toolkit. As of yesterday, I've started at FC17, but in the course of various attempts, I've upgraded to FC18.

First, here's my problem in short: CUDA's repository wants to install for version 319.37, but my current version is 319.32 (the one I can get from rpmfusion). The version of the drivers from rpmfusion works with my video card. When I add CUDA's repository to the list of repos recognized by yum and then try to install from there, the driver won't load claiming version mismatch.

So... I'm stuck. I can't install other CUDA-related libraries, because they list as their dependency something that will not work with my kernel (?) I'm not sure what is that version number.

This is what yum wants to install, once asked to install cuda:

Installing:
 cuda                          x86_64     5.5-22                 cuda     2.4 k
Installing for dependencies:
 cuda-5-5                      x86_64     5.5-22                 cuda     3.3 k
 nvidia-modprobe               x86_64     319.37-1.fc18          cuda      14 k
 xorg-x11-drv-nvidia-devel     i686       1:319.37-2.fc18        cuda     116 k
 xorg-x11-drv-nvidia-devel     x86_64     1:319.37-2.fc18        cuda     116 k
 xorg-x11-drv-nvidia-libs      i686       1:319.37-2.fc18        cuda      28 M
Updating for dependencies:
 xorg-x11-drv-nvidia           x86_64     1:319.37-2.fc18        cuda     5.1 M
 xorg-x11-drv-nvidia-libs      x86_64     1:319.37-2.fc18        cuda      28 M

This is my kernel version:

$ uname -r
3.10.6-100.fc18.x86_64

This is the message I got after trying to install the above (I've reverted to rpmfusion version now).

Aug 16 13:20:59 wvxvvw-laptop avahi-daemon[706]:
  Registering new address record for fe80::5604:a6ff:fe44:3bfa on p5p1.*.
Aug 16 13:20:59 wvxvvw-laptop kernel: [   23.086087] NVRM:
  API mismatch: the client has the version 319.37, but
Aug 16 13:20:59 wvxvvw-laptop kernel: [   23.086087] NVRM:
  this kernel module has the version 319.32.  Please
Aug 16 13:20:59 wvxvvw-laptop kernel: [   23.086087] NVRM:
  make sure that this kernel module and all NVIDIA driver
Aug 16 13:20:59 wvxvvw-laptop kernel: [   23.086087] NVRM:
  components have the same version.
Aug 16 13:20:59 wvxvvw-laptop kdm[1112]:
  X server died during startup

formatted for readability

sgarizvi
  • 16,623
  • 9
  • 64
  • 98

3 Answers3

1

If you are on 64-bit FC18, and you are having trouble with the RPM method, just use the installer method instead.

You can find the CUDA 5.5 kits here.

Here is the direct link to the installer.

Don't forget the getting started guide.

After you've finished with the installer, if it puts a driver version that you don't want on your system, you can go back and re-install your 319.32 driver. Or you can just opt-out of the driver install step when using the installer method.

I would check to see if the cuda 5.5 installed driver works before switching back to 319.32. My sense is that 319.32 should be OK with CUDA 5.5, but you may need to find a newer NVIDIA driver that works with your video card and also works with CUDA. The driver that comes with CUDA 5.5 should work with most or all GPUs. You don't mention which GPU you are using.

And you are getting close votes because questions about system setup are generally off-topic for SO.

Robert Crovella
  • 143,785
  • 11
  • 213
  • 257
  • Hi, sorry, forgot to mention it, the installer fails to complie the kernel module :( I'm getting some GCC error, which probably doesn't point to a correct location. It says something about not ignoring a void variable, but when I open the location it argues about, there doesn't seem to be anything worth ignoring... I'll see if I can opt out of installing it completely, right now I'm trying to install packages one by one with rpm, but I'm only half way through. –  Aug 16 '13 at 13:10
  • As an aside, I checked on Linux & Unix and Superuser - both of them have exactly one question related to CUDA, so I decided to ask it here, not hoping to find the proper audience in those places... –  Aug 16 '13 at 13:11
  • 1
    I suggest you do a clean install of FC18 then. Your FC18 sounds messed up in some non-obvious way. If you don't wish to do that, you can try an opt-out of installing the driver, and just install the cuda toolkit and samples. With the 319.37 driver if that's what you have, it should be OK to have a newer driver. And regarding other forums, there's nothing I can do about their utility. *This* forum is not intended for install type questions, and that is why you are receiving the close votes. – Robert Crovella Aug 16 '13 at 13:26
  • No-no, it's really the NVidia coda vs GCC. They must've used a different complier version, where this warning didn't prevent them from compiling whatever there is in that module. Anyways, looks like I finally got it working by installing only the `cuda-*` rpms with `--nodepts` while keeping the driver from `rpmfusion`. At least I could run a test and didn't get errors. Regardless, thanks for your help / reassuring me that it will probably work with a different driver version too. –  Aug 16 '13 at 13:33
1

add repo:

sudo rpm -Uhv cuda-repo-fedora18-5.5-0.x86_64.rpm 
//you can download it from https://developer.nvidia.com/cuda-downloads

link

install gcc,kernel-devel and update kernel

sudo yum install gcc-c++ kernel-devel
sudo yum update kernel kernel-devel

remove nouveau driver

sudo yum remove xorg-x11-drv-nouveau


sudo yum install nvidia-settings nvidia-kmod

sudo yum install xorg-x11-drv-nvidia //Maybe it is not necessary      
./NVIDIA-Linux-x86_64-xxx.xx.run   //download from http://www.geforce.com/drivers

link

add to /etc/default/grubenter

GRUB_CMDLINE_LINUX_DEFAULT="rdblacklist=nouveau nouveau.modeset=0"

reboot

sudo reboot

install cuda

sudo yum install cuda

add to .bashrc

export CUDA_HOME=/usr/local/cuda-5.5 
export LD_LIBRARY_PATH=${CUDA_HOME}/lib64 

PATH=${CUDA_HOME}/bin:${PATH} 
export PATH
Abesalomi Gogatishvili
  • 193
  • 1
  • 2
  • 8
0

As of 2013-10-02 this appears to be a problem because Fedora 18 has upgraded to Linux kernel version 3.10.x. The nvidia kernel module in the nvidia CUDA repository fails to build in this environment.

There appears to be a similar problem with Fedora 19. Note that Fedora 19 is not in the list of supported distributions for CUDA 5.5. See this conversation in the Nvidia developer forum.

nwhsvc
  • 652
  • 1
  • 6
  • 15
  • If it matters, I'm now using it with rpmfusion drivers (which are a little older then those from cuda.rpm), but I didn't do much testing, only the very basic things, so I don't know if there are any ramifications. –  Oct 02 '13 at 23:40