scripts/train/same_config_w_vicuna_finetune.sh

### !/bin/bash
set -x



nnodes=1
num_gpus=8


DATA_ROOT='/jizhicfs/bojoli'

DATA_PATH=${DATA_ROOT}/llava-next-data/llava-next-data.json
IMAGE_FOLDER=${DATA_ROOT}/llava-next-data/images



VISION_TOWER_NAME='clip-vit-large-patch14-336'
MODEL_NAME='vicuna-7b-1.5'

VISION_TOWER=${DATA_ROOT}/${VISION_TOWER_NAME}
MODEL_PATH=${DATA_ROOT}/${MODEL_NAME}




BASE_RUN_NAME='final_without_mmpe'_${MODEL_NAME}'_'${VISION_TOWER_NAME}

RUN_NAME='final_without_mmpe_finetune'_${MODEL_NAME}'_'${VISION_TOWER_NAME}

deepspeed --num_nodes ${nnodes} --num_gpus ${num_gpus} --master_port=10271 llava/train/train_mem.py \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path ${MODEL_PATH} \
    --version v1 \
    --data_path ${DATA_PATH} \
    --image_folder ${IMAGE_FOLDER} \
    --pretrain_mm_mlp_adapter ./checkpoints/${BASE_RUN_NAME}/mm_projector.bin \
    --unfreeze_mm_vision_tower True \
    --mm_vision_tower_lr 2e-6 \
    --vision_tower ${VISION_TOWER} \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --use_mmpe False \
    --mm_use_im_patch_token False \
    --group_by_modality_length True \
    --image_aspect_ratio anyres \
    --mm_patch_merge_type spatial_unpad \
    --bf16 True \
    --image_grid_pinpoints "[(672,672),(336, 672), (672, 336), (336, 1008),(1008,336),(336,1344),(1344,336)]" \
    --output_dir ./checkpoints/${RUN_NAME} \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 4096 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to none \
    --run_name ${RUN_NAME}