Standardized GPU Driver Helper Functions for AI Backends in LXC Description

[BillyOutlast]

As AI workloads (Ollama, ComfyUI, Lemonade, etc.) have been added to the Proxmox homelab ecosystem, the manual effort required to set up GPU acceleration in LXC remains a significant, and sometimes messy barrier.
Currently, must manually install NVIDIA drivers, configuring ROCm 7.x for different RDNA families, setup Intel oneAPI, and even mess around with all the different implementations of Vulkan drivers.

I am proposing the creation of a standardized GPU helper function or addon within the community-scripts/ProxmoxVE ecosystem to automate these steps.

Why is this useful?

Currently, many scripts (like Ollama or ComfyUI) handle GPU passthrough, and with incoming changes in the ProxmoxVED addressesing the /dev/kfd mapping it seems like a good time to bring in a standardized function or addon to make things more automatic for the end user.

• No post installation configuration required.
• Support Hardware Diversity: Provide a "one-click" way to enable ROCm 7.2+, NVIDIA CUDA, or Intel Arc/OneAPI.
• Simplify AI Deployment: Allow developers to call a single function (e.g., gpu_driver_setup) in their installation scripts.
• Implement Vulkan backends, both open source and properitary ie. Mesa/AMDVLK, vs Nvidia/AMDPRO
• Possible OpenCL

Proposed Technical Implementation

The goal is to move away from fragmented setup guides and into a shared gpu.func or an extension of tools.func that handles:

• NVIDIA CUDA:
  • Detect host driver version.
  • Fetch and install the matching .run installer inside the LXC with the --no-kernel-module flag.
  • Install nvidia-utils and nvidia-container-toolkit.
• AMD ROCm (7.2+):
  • Automate the addition of the Radeon repository.
  • Install amdgpu-install with --usecase=rocm --no-dkms.
  • Set essential environment variables like HSA_OVERRIDE_GFX_VERSION based on detected hardware (RDNA2/3, etc.).
• Intel Arc & iGPU (OneAPI/Vulkan):
  • Install intel-opencl-icd, intel-media-va-driver-non-free, and the oneapi-level-zero stack for AI compute.
  • Standardize Vulkan support via libvulkan1 and vendor ICds.
• Automated Passthrough (Host Side):
  • A script to automate the editing of /etc/pve/lxc/ID.conf with the correct dev0, dev1, and lxc.cgroup2.devices.allow entries for the detected hardware.

Discussion Points

• Should this live as a standalone "GPU-Enabler" script that can be run on any existing LXC, or purely as a shared helper function for other scripts?
• How can we best handle "version pinning" for NVIDIA to ensure the container stays in sync with the host's kernel modules?
• Would there be interest in a pve-gpu-check tool to verify passthrough health across all containers?

I'd love to hear from the maintainers and the community on how to best integrate this into Proxmox community scripts! Thank you

[Builder-DE-TH]

I'm currently running my main local AI loads on bare metal ubuntu 24.04 which is highly sub-optimal for me because of this exact issue. This would be a great help if it happens and I am willing to help test it. I have 3 Mi50 (2 x 32GB and 1 x 16GB) and a RX 580 to test with for older / newly popular AMD cards and multi GPU set-ups using RocM and Vulkan.