when unfreeze_vision_tower set true No matter how the image changes, the model outputs the exact same answer

[Gary2018X]

unfreeze_vision_tower set false Normal output
I want to know how to solve it
the train.sh as follow

#!/bin/bash

MODEL_TYPE=qwen1.5-1.8b
DATA_VERSION=s_all
PRETRAIN_DIR=bunny-$MODEL_TYPE-pretrain
OUTPUT_DIR=bunny-lora-juzao-$DATA_VERSION-7k-uvit-$MODEL_TYPE

mkdir -p ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR

deepspeed bunny/train/train.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed ./script/deepspeed/zero3.json \
    --model_name_or_path models/Qwen1.5-1.8B \
    --model_type $MODEL_TYPE \
    --version bunny \
    --data_path data/$DATA_VERSION/Bunny_all_filtered.json \
    --image_folder  /\
    --vision_tower models/siglip-so400m-patch14-384 \
    --pretrain_mm_mlp_adapter models/bunny-pretrain-qwen1.5-1.8b-siglip/mm_projector.bin \
    --mm_projector_type mlp2x_gelu \
    --image_aspect_ratio pad \
    --group_by_modality_length False \
    --bf16 True \
    --output_dir ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR \
    --num_train_epochs 3 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --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 \
    --unfreeze_vision_tower True\
    --report_to none | tee 2>&1 ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR/log.txt

[Isaachhh]

What is the "exact same" answer? Maybe overfitting?

[Gary2018X]

My answers are diverse
I have also trained a few times
I found that the output randomly selected one of my answers
and My previous code used this training script normally
I updated to the latest code and there was a problem

[Isaachhh]

The recent commit 08273ac is due to #130.