for ((rank_id=0;rank_id<$NUM_RANKS;rank_id++)); do
    device_id=${RANK_DEVICE_ID[${rank_id}]}
    numa_node=${DEVICE_NUMA_NODE[${device_id}]}
    if [ "$rank_id" == "0" ]; then
        numactl --cpunodebind=${numa_node} --membind ${numa_node} \
            python -m paddle.distributed.launch \
                --log_dir log/worker_${rank_id} \
                --master 127.0.0.1:${NUMA_PORT} \
                --rank ${rank_id} \
                --devices ${device_id} \
                --nnodes ${NUM_RANKS} \
                ${infer_cmd} > /opt/worker_$rank_id.log 2>&1 &
    else
        numactl --cpunodebind=${numa_node} --membind ${numa_node} \
            python -m paddle.distributed.launch \
                --log_dir log/worker_${rank_id} \
                --master 127.0.0.1:${NUMA_PORT} \
                --rank ${rank_id} \
                --devices ${device_id} \
                --nnodes ${NUM_RANKS} \
                ${infer_cmd} > /opt/worker_$rank_id.log 2>&1 &
    fi
    task_pids[${rank_id}]=$!
    echo "Rank ID: $rank_id, Device ID: $device_id, bind to NUMA node: ${numa_node}, PID: ${task_pids[${rank_id}]}"
done