← Back to Blog

vLLM stays Pending because the GPU resource name does not match


Idle inference server with a GPU and no workload lights

The vLLM pod stays Pending with Insufficient nvidia.com/gpu. The node has a card. I can see it. The request and the advertisement are different resource names, so to the scheduler the node has nothing this pod can take.

The problem

A full card is nvidia.com/gpu. A MIG slice is a longer name, like nvidia.com/mig-1g.10gb. Time-slicing still uses nvidia.com/gpu, but the count is the number of replicas, not the number of cards. Copying a manifest from another cluster keeps the old name.

The check

kubectl describe node <gpu-node> | sed -n '/Allocatable/,/Events/p'
kubectl describe pod <vllm> | grep -A2 'Insufficient\|nvidia.com'

Read the allocatable keys. Then make the pod request that key, with a count the node actually has.

resources:
  limits:
    nvidia.com/gpu: 1

If the node only lists a MIG name, the limit has to be that name. Raising replicas does not change the name. A second pod with the same wrong request stays Pending too.

The setup is in the operator note.

Commercial Engineering & Platform Infrastructure Advisory

Need hands-on senior engineering to audit, scale, or stabilize your infrastructure?

From bare-metal Kubernetes and immutable Talos Linux to drastically cutting cloud infrastructure bills, resolving PostgreSQL deadlocks, and architecting sovereign AI systems, I work directly with production engineering teams to deliver resilient foundations.