How to use nccl for communication when using multi-node GPU for distributed training, and use the ib card on the machine？