LLaVA: [Question] OOM on finetuning vicuna-7b llava model on 4*A800 80G, anything wrong with my cfg?

Question

Thanks for the great work~

Also, it looks like A800 cannot enable flash-attn. (error screenshot below)

python \
    llava/train/train.py \
    --model_name_or_path /root/devroot/models/vicuna-7b-v1.3 \
    --version v1 \
    --data_path /root/devroot/datasets/llava_instruct/llava_instruct_150k.json \
    --image_folder /root/devroot/datasets/llava_instruct/coco-train2017/ \
    --vision_tower openai/clip-vit-large-patch14 \
    --pretrain_mm_mlp_adapter ./checkpoints/vicuna-7b-pretrain/mm_projector.bin \
    --mm_vision_select_layer -1 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --bf16 True \
    --output_dir ./checkpoints/vicuna-7b-finetune \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 32 \
    --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 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb