[elrepo] nVidia kmods and CUDA

[Phil Perry]

On 19/08/15 18:50, Phil Perry wrote:
> On 19/08/15 18:26, Pat Riehecky wrote:
>> I'm running into an unexpected dependency problem loading the CUDA rpms
>> and the elrepo kmod:
>>
>> # yum install cuda
>> Loaded plugins: nvidia, security
>> Setting up Install Process
>> Resolving Dependencies
>> --> Running transaction check
>> ---> Package cuda.x86_64 0:7.0-28 will be installed
>> --> Processing Dependency: cuda-7-0 = 7.0-28 for package:
>> cuda-7.0-28.x86_64
>> --> Running transaction check
>> ---> Package cuda-7-0.x86_64 0:7.0-28 will be installed
>> --> Processing Dependency: cuda-runtime-7-0 = 7.0-28 for package:
>> cuda-7-0-7.0-28.x86_64
>> --> Running transaction check
>> ---> Package cuda-runtime-7-0.x86_64 0:7.0-28 will be installed
>> --> Processing Dependency: cuda-drivers >= 346.46 for package:
>> cuda-runtime-7-0-7.0-28.x86_64
>> --> Running transaction check
>> ---> Package cuda-drivers.x86_64 0:346.46-0 will be installed
>> --> Processing Dependency: nvidia-kmod >= 1:346.46 for package:
>> cuda-drivers-346.46-0.x86_64
>> --> Processing Dependency: xorg-x11-drv-nvidia-gl(x86-64) >= 1:346.46
>> for package: cuda-drivers-346.46-0.x86_64
>> --> Processing Dependency: nvidia-uvm-kmod >= 1:346.46 for package:
>> cuda-drivers-346.46-0.x86_64
>> --> Processing Dependency: xorg-x11-drv-nvidia-devel(x86-64) >= 1:346.46
>> for package: cuda-drivers-346.46-0.x86_64
>> --> Processing Dependency: xorg-x11-drv-nvidia-libs(x86-64) >= 1:346.46
>> for package: cuda-drivers-346.46-0.x86_64
>> --> Processing Dependency: xorg-x11-drv-nvidia >= 1:346.46 for package:
>> cuda-drivers-346.46-0.x86_64
>> --> Running transaction check
>> ---> Package nvidia-kmod.x86_64 1:346.46-2.el6 will be installed
>> ---> Package nvidia-uvm-kmod.x86_64 1:346.46-3.el6 will be installed
>> ---> Package xorg-x11-drv-nvidia.x86_64 1:346.46-1.el6 will be installed
>> ---> Package xorg-x11-drv-nvidia-devel.x86_64 1:346.46-1.el6 will be
>> installed
>> ---> Package xorg-x11-drv-nvidia-gl.x86_64 1:346.46-1.el6 will be installed
>> ---> Package xorg-x11-drv-nvidia-libs.x86_64 1:346.46-1.el6 will be
>> installed
>> --> Processing Conflict: nvidia-x11-drv-352.30-1.el6.elrepo.x86_64
>> conflicts xorg-x11-drv-nvidia
>> --> Finished Dependency Resolution
>> Error: nvidia-x11-drv conflicts with
>> 1:xorg-x11-drv-nvidia-346.46-1.el6.x86_64
>>  You could try using --skip-broken to work around the problem
>>  You could try running: rpm -Va --nofiles --nodigest
>> # rpm -qa |grep nvidia
>> yum-plugin-nvidia-1.0.2-1.el6.elrepo.noarch
>> kmod-nvidia-352.30-1.el6.elrepo.x86_64
>> nvidia-x11-drv-352.30-1.el6.elrepo.x86_64
>> nvidia-detect-352.30-1.el6.elrepo.x86_64
>>
>> I suspect some new "provides" might be required.
>>
>> Pat
>>
> 
> Hi Pat,
> 
> I'm not overly familiar with anything CUDA, but from the above
> transaction it does look like we might be able to get away with adding a
> provides for cuda-drivers to our nvidia package(s), which should then
> prevent yum trying to pull in their nvidia drivers to meet that
> dependency requirement.
> 
> As a workaround, you could try manually installing the cuda and
> cuda-runtime packages from above with --nodeps and seeing if that works
> for you (there may be other unresolved deps though).
> 
> In the meantime, I will try to get you a set of updated packages built
> with the extra provides and into the testing repo for you to test.
> 
> If there are any CUDA users out there, please feel free to chip in :-)
> 
> Phil
> 
> 

Rather than build a whole new set of nvidia packages just to add missing
provides, I've just built a dummy cuda-drivers package with the missing
provides for now (this seems to be how nvidia does it):

$ rpm -qlp --provides cuda-drivers-352.30-1.el6.elrepo.x86_64.rpm
cuda-driver = 352.30
cuda-drivers = 352.30
cuda-drivers = 352.30-1.el6.elrepo
cuda-drivers(x86-64) = 352.30-1.el6.elrepo
(contains no files)

I've pushed it to the el6 testing repo so it should start syncing to
mirrors shortly.

Please try:

yum --disablerepo=\* --enablerepo=elrepo-testing install cuda-drivers

(make sure you are installing cuda-drivers from elrepo-testing)

and then try:

yum install cuda

and see if yum now completes the transaction cleanly.

Assuming this works, I'll then add the missing provides to the next
nvidia drivers release.

One question, I note from this cuda repo:

http://developer.download.nvidia.com/compute/cuda/repos/rhel6/

that only x86-64 cuda packages are available, so would I be correct in
assuming cuda is not available on x86, so I should only add cuda-driver
provides to 64-bit packages?

Likewise, does it make any sense to add to older legacy drivers (340.xx,
304.xx)?

Thanks,

Phil