Hello!
New user here, happy to be a part of this community!
I have been trying to set up CUDA on an OCI instance running an Oracle UEK 5.15.0 kernel. I followed the guidance here: CUDA Installation Guide for Linux (nvidia.com), but at the step when I have to install the kernel-devel and kernel-header:
sudo dnf install kernel-devel-$(uname -r) kernel-headers-$(uname -r)
I ran into the issue that no matches are found for arguments:
kernel-devel-5.15.0-204.147.6.2.el8uek.x86_64
kernel-headers-5.15.0-204.147.6.2.el8uek.x86_64
Here, for the kernel-devel-5.15 I was able to find a package in an Oracle repo, however, the kernel-headers are no longer provided as part of Oracle Linux.
I installed only the kernel-devel and proceeded with the install, however after the reboot I got the infamous “NVIDIA-SMI has failed because it couldn’t communicate with the NVIDIA driver” when running nvidia-smi.
I decided to check if this has anything to do with the kernel devel/header and set the default Linux kernel to RHEL8 4.18, for which there are easily available kernel and devel packages. After doing a fresh install with the default kernel being 4.18 - nvidia-smi worked and torch.cuda.available_devices was working as expected. However, I can’t work on an old kernel, as torch/transformers have peculiarities with old versions of linux kernels.
My question - what would be the optimal approach to set up Cuda on an Oracle UEK 5.15 kernel? Thank you!