distributed_data_parallel_train.sh

#!/usr/bin/env bash
export NCCL_P2P_DISABLE=1
# export NCCL_DEBUG=INFO
export NCCL_IB_DISABLE=1
export NCCL_SOCKET_IFNAME=enp97s0f0


PYTHON=${PYTHON:-"python"}
CONFIG=$1
NUM_GPUS=$2
GPUS=$3
PORT=${PORT:-29500}


CUDA_VISIBLE_DEVICES=$GPUS $PYTHON -m torch.distributed.launch --nproc_per_node=$NUM_GPUS --master_port=$PORT \
        main/train.py --cfg $CONFIG ${@:4}