#!/bin/bash
lr=2e-5

output_dir="Qwen2-Audio-7B-Instruct-train-all-full-new"
if [ ! -d "./output/${output_dir}" ]; then
  mkdir -p "./output/${output_dir}"
fi

cp $0 "./output/${output_dir}"

deepspeed --master_port=1222 --include localhost:0,1,2,3,4,5,6,7 src/train.py \
    --deepspeed "/GLOBALFS/gznwp_3/qxj/liangjh/LLaMA-Factory-audio/examples/deepseed/ds_z2_config.json" \
    --stage sft \
    --model_name_or_path "/GLOBALFS/gznwp_3/qxj/liangjh/model_set/qwen/Qwen2-Audio-7B-Instruct" \
    --do_train \
    --dataset audio_deepfake_val_new_2w,audio_emotion_train_1w5_wo_happy,audio_speaker_recognition_random_order_train \
    --freeze_vision_tower \
    --preprocessing_num_workers 60 \
    --ddp_timeout 180000000 \
    --cutoff_len 3072 \
    --template qwen2_audio \
    --finetuning_type full \
    --output_dir "./output/${output_dir}" \
    --overwrite_cache \
    --overwrite_output_dir \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 1 \
    --save_strategy="epoch" \
    --save_only_model True \
    --num_train_epochs 3 \
    --report_to "tensorboard" \
    --learning_rate ${lr} \
    --warmup_ratio 0.05 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --plot_loss \
    --fp16 \
    --log_level info \
    2>&1 | tee "./output/${output_dir}/${output_dir}.log" &


wait