in my case, I found that a few dead pops were taking up the GPU resources on a few worker nodes which didn't allow for scheduling. I would suggest check your pod resources and limits.
kubectl describe node NODENAME
Check to see if anything is maxed out or resources that are expected are not avaialble.