### !/bin/bash
set -x

nnodes=1
num_gpus=8

DATA_ROOT='/jizhicfs/bojoli'

DATA_PATH=${DATA_ROOT}/llava-pretrain/blip_laion_cc_sbu_558k.json
IMAGE_FOLDER=${DATA_ROOT}/llava-pretrain/images



VISION_TOWER_NAME='siglip'
MODEL_NAME='qwen2.5-0.5b-i'

VISION_TOWER=${DATA_ROOT}/${VISION_TOWER_NAME}
MODEL_PATH=${DATA_ROOT}/${MODEL_NAME}
RUN_NAME='mmpe'_${MODEL_NAME}'_'${VISION_TOWER_NAME}

deepspeed --num_nodes ${nnodes} --num_gpus ${num_gpus} --master_port=10270 llava/train/train_mem.py \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ${MODEL_PATH} \
    --version plain \
    --data_path ${DATA_PATH} \
    --image_folder ${IMAGE_FOLDER} \
    --vision_tower ${VISION_TOWER} \
    --mm_projector_type mlp2x_gelu \
    --tune_mm_mlp_adapter True \
    --unfreeze_mm_vision_tower False \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --mm_patch_merge_type spatial_unpad \
    --image_aspect_ratio anyres \
    --group_by_modality_length False \
    --bf16 True \
    --output_dir ./checkpoints/${RUN_NAME} \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --image_grid_pinpoints "[(384, 768), (768, 384), (768, 768), (1152, 384), (384, 1152)]" \
    --use_mmpe True \
    --save_strategy "steps" \
    --save_steps 400 \
    --save_total_limit 1 \
    --learning_rate 1e-3 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 32768 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to none \
    --run_name ${RUN_NAME}