Qwen2.5-VL Expect all tensors on the same devices

[yyy195]

Dear author,
when I evaluate Qwen2.5-VL-3B on only one GPU and set num_processes=1, the error occurred!
"Error during evaluation: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select). Please set --verbosity=DEBUG to get more information."

I've searched issues for this , but all are about multiple GPUS, so could you help me solve this. ‼️I've tried "device_map=auto",but it not works
My .sh file is as follows

export HF_HOME="~/.cache/huggingface"
export CUDA_VISIBLE_DEVICES=0

accelerate launch --num_processes=1 --main_process_port=12344 -m lmms_eval
--model qwen2_5_vl
--model_args=pretrained=./datasets/Qwen2.5-VL-3B-Instruct,max_pixels=12845056,attn_implementation=flash_attention_2,interleave_visuals=False,device_map=auto
--tasks mathvision_test
--device cuda:0
--output_path ./logs/
--batch_size 4