You are not logged in.
Gentlemen!
I am running the NVIDIA legacydriver (nvidia-470xx) as my card is not supported by the current nvidia driver. Now I would like to have a try on cuda with tensorflow (or vice versa, if it pleases you) but I am running into compatiblity issues as the current cuda toolkit (cuda-12.6.3) does require a more recent driver to work with.
Now I found that NVIDIA supplies compatibility packages for new cuda toolkits to work with older drivers ... they call them cuda-compat-12.6 and the like ... they are said to reside in the "repositories" but I only managed to spot some for non-arch distributions ...
Having said all that, here comes my question: How to install compatibility packages for cuda under arch linux? Or would it be better to use user level tools like micromamba to achieve that (as suggested here).
Any suggestion is appreciated.
Cheers
Last edited by DrJ (2025-01-29 15:30:50)
Offline
If you know what the latest cuda version is that does support your card & driver, you can use that one from aur .
cuda-12.2, 11.7, 11.1, 10.2 , 10.0, 9.2, 9.0 and 8.0 are all present .
Can you post a link to info about those compat packages ?
Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.
clean chroot building not flexible enough ?
Try clean chroot manager by graysky
Offline
Mod note: moving to AUR Issues.
Sakura:-
Mobo: MSI MAG X570S TORPEDO MAX // Processor: AMD Ryzen 9 5950X @4.9GHz // GFX: AMD Radeon RX 5700 XT // RAM: 32GB (4x 8GB) Corsair DDR4 (@ 3000MHz) // Storage: 1x 3TB HDD, 6x 1TB SSD, 2x 120GB SSD, 1x 275GB M2 SSD
Making lemonade from lemons since 2015.
Offline
If you know what the latest cuda version is that does support your card & driver, you can use that one from aur .
cuda-12.2, 11.7, 11.1, 10.2 , 10.0, 9.2, 9.0 and 8.0 are all present .
Can you post a link to info about those compat packages ?
Hello!
Thank you for your reply! The last cuda version that is compatible with 470xxx is 11.4 ... which is available in AUR ... have not tested yet.
For the compatibility packages please be as kind as referring to
https://docs.nvidia.com/deploy/cuda-com … patibility
especially the section "forward compatibility"
With forward compatibility they mean running a recent cuda toolkit with an older kernel driver by changing the user space driver (i.e. libcuda.so) accordingly.
Cheers
Offline
Mod note: moving to AUR Issues.
Well moving this to AUR is not really helpful, as a compatibility packages helps the _current_ cuda (arch) packages to be used with a legacy driver (which is in deed an AUR package).
Offline
Looking at the nvidia doc, the compat package needs to be tailored to the specific driver and that may require multiple packages.
AUR seems best solution for that.
Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.
clean chroot building not flexible enough ?
Try clean chroot manager by graysky
Offline
Looking at the nvidia doc, the compat package needs to be tailored to the specific driver and that may require multiple packages.
AUR seems best solution for that.
Well ... me, IMHO, do not think so ... but this is not my forum anyway ...
Let's get back to the question I posed: Does anybody have any experience in using the compat package approach? ... and is willing to share this?
Cheers
Offline
python-tensorflow is currently broken in the arch Repo's so if you are going to have to use a venv for that anyway I would suggest using a version built for the CUDA version your driver natively supports. From section 5 of the document linked the GPUs supported are "11.4 UMD (User Mode Driver) and later will extend forward compatibility support to select NGC Ready NVIDIA RTX boards. Prior to that forward compatibility will be supported only on NVIDIA Data Center cards." does that cover your GPU?
Offline
python-tensorflow is currently broken in the arch Repo's so if you are going to have to use a venv for that anyway I would suggest using a version built for the CUDA version your driver natively supports. From section 5 of the document linked the GPUs supported are "11.4 UMD (User Mode Driver) and later will extend forward compatibility support to select NGC Ready NVIDIA RTX boards. Prior to that forward compatibility will be supported only on NVIDIA Data Center cards." does that cover your GPU?
Thank you very much for that piece of information. I will follow your advice and use that approach. It is micromamba in my case, but specific versions of cuda can be selected there as well.
Concerning the support of my graphics card I must admit that I am not sure if this the case. I do not really understand nvidia's naming scheme and competing naming conventions... the 470xx driver should work with the 11.4 UMD according to the documentatio ... but yes, I am aware of potential problems there.
Thank you very much for your reply!
Offline
Well to finish it up: I could not figure out how to get hold of the appropiate compat packages and could not find out how to install them.
On the upside there is a workaround - as hinted above - in installing a toolkit package from AUR that matches the driver.
It took a complete night to install it, as a specific gcc version is built with it. Nevertheless I could compile some cuda sample files with nvcc ... it seems to work for C.
Let me emphasize that I did not thoroughly test it, as I am actually after a Python based solution ...
Let me finish this off by thanking all the contributers for their helpful remarks.
Cheers
Offline