[Question] LLaVA 1.5 7B model fine-tune -- pydantic

[yiwei-chenn]

Question

When I use my own pre-trained mlp adapter to finetune the LLaVA 1.5 7B model, I use the finetune_lora.sh like

deepspeed llava/train/train_mem.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed ./scripts/zero3.json \
    --model_name_or_path lmsys/vicuna-7b-v1.5 \
    --version v1 \
    --data_path ./my_own.json \
    --image_folder ./llava-finetune \
    --vision_tower openai/clip-vit-large-patch14-336 \
    --pretrain_mm_mlp_adapter ./checkpoints/llava-v1.5-7b-pretrain/mm_projector.bin \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ./checkpoints/llava-v1.5-7b-lora \
    --num_train_epochs 1 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 1 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 50000 \
    --save_total_limit 1 \
    --learning_rate 2e-4 \
    --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

And I face the problem like

This problem was caused by

'pydantic_core._pydantic_core.ValidationError: 1 validation error for DeepSpeedZeroConfig
stage3_prefetch_bucket_size
Input should be a valid integer, got a number with a fractional part [type=int_from_float, input_value=15099494.4, input_type=float]'

I used the same procedure to use LoRA fine-tune the LLaVA 1.5 13B version, but did not cause the same problem.

Does anyone know how to solve that?

[yiwei-chenn]

After checking deepspeedai/DeepSpeed#6525, the problem seems caused by the mismatch between transformers and deepspeed.

I solved the problem by downgrading deepspeed:

pip install deepspeed==0.14.5