This issue is caused by the NVIDIA vGPU scheduling policy. On many NVIDIA GRID/vGPU architectures (prior to latest MIG capabilities), you cannot mix different vGPU profile sizes (e.g., 2Q and 4Q) on the same physical GPU.

Current State (Breadth Mode): The scheduler is likely distributing the "2Q" VMs evenly across all available GPUs. This means every physical GPU has at least one "2Q" profile active, "tainting" the GPU so it can only accept other "2Q" VMs.

The Fix (Depth Mode): By changing the allocation policy to Depth-First, the scheduler will fill up the first physical GPU with "2Q" VMs until it is full before moving to the next. This consolidates the "2Q" workloads onto fewer cards, leaving other physical GPUs completely empty. These empty GPUs can then accept the "4Q" profile type, allowing both profiles to co-exist on the same host (though on different physical cards).