On 18/11/2025 13:31, KK CHN wrote:
> List,
>
> I am having NVIDIA L40s, PCle  GPU installed in one of Ubuntu Server
> Box,  How can I make use this  NVIDIA L40s card to support multiple
> users for running their AI work loads.
>
> How to achieve this virtualization support for the NVIDIA Card.
nova support both time sliced and hardware patitioned (MIG) mode for
nvida gpus
https://docs.openstack.org/nova/latest/admin/virtual-gpu.html

and it support there new way of doign MIG without mdevs via vfio
vartiant driver via sriov/pci passthough
https://docs.openstack.org/nova/latest/admin/pci-passthrough.html
https://specs.openstack.org/openstack/nova-specs/specs/2025.1/implemented/enable-vfio-devices-with-kernel-variant-drivers.html

>
> Some body can shed some light on this regard , what are the steps need
> to perform to get the GPU card virtualized ?
nova has the most advanced supprot for this usecase today including live
migration and in the case of the varaint driver approch numa affinity
i am currently proposing completing the supprot for cyborg for mdev
based gpus virutalisation
https://review.opendev.org/c/openstack/nova-specs/+/967515
but today you cant do gpu virtualistion with cyborg as nova will not
accept the mdev type as a mode of operation.
using cyborg also would not supprot numa affinity or live migration so
today novas native supprot for gpus either in fully passthough, time
sliced or hardware pattioned
is much mroe advanced then cybrog.

>
> Thanks in advance,
> Krishane