MIG partitions GPUs physically. Time-Slicing takes turns. MPS runs kernels in parallel. When to use each GPU sharing strategy on Kubernetes.