train.sh

#!/bin/bash
NUM_PROC=$1
shift
# the localhost fixes are from https://github.com/pytorch/pytorch/issues/73320
torchrun  --rdzv_backend c10d --rdzv_endpoint localhost:0 --nproc_per_node=$NUM_PROC train.py "$@"