Is "karpenter.k8s.aws/instance-gpu-memory" per-gpu or per-node?

[zevisert-aaico]

Looking through https://karpenter.sh/docs/reference/instance-types, the values for karpenter.k8s.aws/instance-gpu-memory seem inconsistent.

For instance:

• Evidence of it being per-device:
  • g4dn.8xlarge has 1x NVIDIA T4 GPU and lists instance-gpu-memory = 16384
  • g4dn.12xlarge has 4x NVIDIA T4 GPU and lists instance-gpu-memory = 16384
• Evidence of it being per-node:
  • g6e.16xlarge has 1x NVIDIA L40S GPU and lists instance-gpu-memory = 45776
  • g6e.24xlarge has 4x NVIDIA L40S GPU and lists instance-gpu-memory = 183105
  • g6e.48xlarge has 8x NVIDIA L40S GPU and lists instance-gpu-memory = 366211

Does anyone know if whether the karpenter.k8s.aws/instance-gpu-memory label is per-gpu or for all GPUs available on the node?