LLaMA-Factory-310P3/run_once.sh

#!/bin/bash


# bash run_once.sh lora_sft Qwen-7B 8 50
# bash run_once.sh lora_sft Qwen-7B 1 50
# bash run_once.sh lora_sft 9g-8B 8 500
# bash run_once.sh inference Qwen-7B 8 50
run_type="$1"
model="$2"
gpu_cnt="$3"
max_steps="$4"

echo "run_type: ${run_type}"
echo "model: ${model}"
echo "gpu_cnt: ${gpu_cnt}"
echo "max_steps: ${max_steps}"

current_datetime=$(date +%Y%m%d%H%M%S)

source /usr/local/Ascend/ascend-toolkit/set_env.sh

if [ "${run_type}"="lora_sft" ]; then
    run_name="${run_type}_${model}_${gpu_cnt}_gpu_${max_steps}_step_${current_datetime}"
    
else
    run_name="${run_type}_${model}_${gpu_cnt}_gpu_${current_datetime}"
fi

output_dir="./results/910b/${run_name}"

if [ ! -d "$output_dir" ]; then
    mkdir -p "$output_dir"
    echo "output_dir created: $output_dir"
else
    echo "output_dir exists: $output_dir"
fi

# echo "${run_type} ${model} ${gpu_cnt} ${max_steps} ${run_name} ${output_dir}"
python prepare_yaml_file.py ${run_type} ${model} ${max_steps} ${run_name} ${output_dir}

export USE_MODELSCOPE_HUB=1

echo "Start recording npu status "
bash npu_status.sh ${output_dir} 60 0 &
npu_status_pid=$!
# echo "${npu_status_pid}"

if [ "${gpu_cnt}" = "1" ]; then
   ASCEND_RT_VISIBLE_DEVICES=0 llamafactory-cli train ${output_dir}/${run_name}.yaml 2>&1 | tee -i "${output_dir}/log.txt" &
   train_pid=$!
   echo "Start single npu train"
else
   FORCE_TORCHRUN=1 llamafactory-cli train ${output_dir}/${run_name}.yaml 2>&1 | tee -i "${output_dir}/log.txt" &
   train_pid=$!
   echo "Start multi npu train"
fi

wait $train_pid
train_status=$?
echo "Train ended"

sleep 60
kill $npu_status_pid
echo "Npu status ended"

echo "train_status： ${train_status}"
if [ $train_status -ne 0 ]; then
    output_dir="${output_dir}_fail"
fi

python3 -c "import moxing as mox; mox.file.copy_parallel('${output_dir}', 'obs://xty/results/${run_name}')"
echo "${run_name} uploaded"
rm -r ${output_dir}