In KVM (Ubuntu 22.04), we are running an A100 PCIe 40GB in MIG mode.
GPU partitions are allocated as mdev links to virtual machines and are partitioned to the smallest possible size (1g.5g).
Now, end users want virtual machines (VMs) to compete for every GPU resource.
Some virtual machines occasionally require more GPU time.
My understanding is that this cannot be accomplished in MIG mode due to the fixed partitions.
I’m grateful