The --cuda-mpi flag denotes GPU usage and shouldn't require the --gpu flag

[adamlavely]

The --cuda-mpi flag is utilized by users who wish to use CUDA aware MPI. This inherently requires GPU access. Currently the --gpu flag is also required, but the functionality added there should be included in the --cuda-mpi flag.

[pwxy]

For backward compatibility, it would be best to accept both "--cuda-mpi" and "--cuda-mpi --gpu" for GPU-Aware MPI.