train_rm.sh

set_n_least_used_CUDA_VISIBLE_DEVICES() {
    local n=${1:-"9999"}
    echo "GPU Memory Usage:"
    local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv \
        | tail -n +2 \
        | nl -v 0 \
        | tee /dev/tty \
        | sort -g -k 2 \
        | awk '{print $1}' \
        | head -n $n)
    export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g')
    echo "Now CUDA_VISIBLE_DEVICES is set to:"
    echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
}

set_n_least_used_CUDA_VISIBLE_DEVICES 2

torchrun --standalone --nproc_per_node=2 train_reward_model.py \
   --pretrain  <your pretrain path> \
   --model 'bloom' \
   --strategy colossalai_zero2 \
   --loss_fn 'log_sig'\
   --save_path <your model saving path>\
   --dataset 'Anthropic/hh-rlhf'\