The passthrough of vGPUs using the built-in Proxmox VE functionality is quite restricted in terms of utilization of NVIDIA features.
An example of that would be inability to set vgpu_params without installing vgpu_unlock, even though it is not necessarily needed for anything else, if you're using a GPU that is natively supported by the vGPU driver.
I created a hook script that is supposed to address the problem by using native QEMU passthrough instead of the simplified Proxmox VE one, which then allows to set vgpu_params before the mediated device is created. Proxmox VE does not allows to do that, and it is also not possible to set them once the mediated device is already created.
I suggest integrating nvidia-vgpu-lifecycle.pl as an installation step to this script, perhaps optional, to allow for seamless integration with the VMs.
The passthrough of vGPUs using the built-in Proxmox VE functionality is quite restricted in terms of utilization of NVIDIA features.
An example of that would be inability to set
vgpu_params
without installingvgpu_unlock
, even though it is not necessarily needed for anything else, if you're using a GPU that is natively supported by the vGPU driver.I created a hook script that is supposed to address the problem by using native QEMU passthrough instead of the simplified Proxmox VE one, which then allows to set
vgpu_params
before the mediated device is created. Proxmox VE does not allows to do that, and it is also not possible to set them once the mediated device is already created.I suggest integrating
nvidia-vgpu-lifecycle.pl
as an installation step to this script, perhaps optional, to allow for seamless integration with the VMs.Please see https://gist.github.com/Serpentiel/212b576e2b362ce2ca6b063398973ad4 for more details. Currently its usage is limited to 1 vGPU only, this might change in future if there is demand for it.