step5_train_pt_multi_gpu.sh

#!/bin/bash

export NCCL_P2P_DISABLE="1"
export NCCL_IB_DISABLE="1"

NPROC_PER_NODE=4
NNODES=1
RANK=0
MASTER_ADDR=127.0.0.1
MASTER_PORT=29500

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun \
    --nproc_per_node $NPROC_PER_NODE \
    --nnodes $NNODES \
    --node_rank $RANK \
    --master_addr $MASTER_ADDR \
    --master_port $MASTER_PORT \
    shell_train.py train_config/step5_train_pt.yaml