MIG with Docker and Kubernetes

Docker:

# Run a container on a specific MIG instance (use UUID from nvidia-smi -L)
docker run --gpus '"device=MIG-GPU-abc123/0/0"' nvcr.io/nvidia/cuda:latest

Kubernetes (requires NVIDIA GPU Operator + MIG Manager):

# Pod requesting a MIG instance
resources:
  limits:
    nvidia.com/mig-1g.5gb: 1

MIG strategies in Kubernetes:

• single — All MIG instances on a node are the same size
• mixed — Different sized instances allowed

Fleet Command: MIG can be configured from the Fleet Command UI without SSH access.