OOM problem when fine-tune reward model with LLaMA in step 2

cd training/step2_reward_model_finetuning/
bash training_scripts/single_node/run_llama.sh

run_llama.sh contains

#!/bin/bash
# Copyright (c) Microsoft Corporation.
# SPDX-License-Identifier: Apache-2.0

# DeepSpeed Team
OUTPUT=$1
ZERO_STAGE=$2
if [ "$OUTPUT" == "" ]; then
    OUTPUT=./output
fi
if [ "$ZERO_STAGE" == "" ]; then
    ZERO_STAGE=0
fi
mkdir -p $OUTPUT

deepspeed main.py \
   --data_path some_data \
   --data_split 2,4,4 \
   --model_name_or_path path_to_llama \
   --num_padding_at_beginning 1 \
   --per_device_train_batch_size 1 \
   --per_device_eval_batch_size 1 \
   --gradient_checkpointing \
   --max_seq_len 512 \
   --learning_rate 5e-5 \
   --weight_decay 0.1 \
   --num_train_epochs 1 \
   --disable_dropout \
   --gradient_accumulation_steps 1 \
   --lr_scheduler_type cosine \
   --num_warmup_steps 0 \
   --seed 1234 \
   --zero_stage $ZERO_STAGE \
   --deepspeed \
   --output_dir $OUTPUT \
   &> $OUTPUT/training.log

Even if I set the per_device_train_batch_size = 1 and use gradient_checkpointing, I still have an OOM problem. Any solutions?

Reproduce the OOM from training/step2_reward_model_finetuning/ using training_scripts/single_node/run_llama.sh, then inspect main.py and the script's DeepSpeed settings. Compare the batch, sequence length, gradient checkpointing, and zero-stage options shown in the issue; the payload does not specify hardware or a confirmed fix, so completion is not clearly defined.