Weird Evaluation Output when using Qwen2.5-VL-7B-Instruct

Hi,

Thanks for your contribution to the community for bringing this utility!

I am running into a problem when inference with Qwen2.5-VL models. The outputs for MMStar are as follows,

Infer Qwen2.5-VL-7B-Instruct/MMStar, Rank 2/4:   0%|                                                                                                                                                       | 0/375 [00:00<?, ?it/s]
 A A 144441441111144141454445144414542444224441414241424242424441414244142414241414441414442414141414144441414144414144414141414141414141411414124412444412444124412144412141214412141412141141211412114121141211421211211214412111
41211121112144121111211114412111114114121111121111211111441211112111121111211111411112111121111211112111121111211112111121111211112111114411111111211111111111111111111111111111111111111111111121111111111212112111111112212212212
21221221221221221221221221111221111221111221111221111212212212212122122122112212112121212212121212122121212121212121221211212112121212121212121212121212121211441211111111112122122121121121212121212114212212121212122121212121212
21212112121212122221221211212121221212121121212121221212112121121121212122121212112121121212112121121121212212212112121212122221221211212121221221211211212121221221211211211212212121212122121221211212121221212212112112121212212
21211211212112121212122121211211212112121212121221212112212112112121212212112112112112122121221211211212122122122122122121221121221121212121221221212121221212212112121212212212121212212212122122121221212211221211212121212212212
1122112122121122112121212122212212112211212212122121221212212122122121221221211221121221121221121212121221221211221121212121221221211221121221121221121212121221121211212112121212122222122122121121211211212212112211212112122212112212112112212112122121221221211221121212122122122121221122121121211212211212112112121221221221211211211212122121121121211212212211212211212112112122112112112112122222222222222221222121121122121221221211221121221121221212211212
21121221121221121212212122112211212121221221221211222122112211212211212211212211212212212211222121221122112122112122112122112122112122112122112122112112221212211221121221122212122122121122112212112211212211212211212211212211212
21121221121221121221121221121221121221121221121221151221221221221221221221221121221121221121221121122212212212211221121221122212122121121221121221121121221121221121121212221221221221112222121221212211221211212222222222222222212
2222222                                                                                                                                                                                                                            
Infer Qwen2.5-VL-7B-Instruct/MMStar, Rank 1/4:   0%|▍                                                                                                                                            | 1/375 [00:44<4:37:02, 44.44s/it]
 A A A A  A  A  A  A  A  A  A  A  44  11  11  11  11  11  11  11  11  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1
  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  
1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1 
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1
  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  
1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1 
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1 
 1  1 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1 
 1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1  1
  1  1  1  1  1                        

I have tried other models such as InternVL3-8B and works well. Any suggestions on solution to this? Thanks.

Reproduce the reported Qwen2.5-VL-7B-Instruct evaluation on MMStar and compare its output with InternVL3-8B, which reportedly works. Trace the inference path used by the evaluation toolkit and identify why Qwen2.5-VL produces repeated token-like output; done means explaining the discrepancy and verifying a corrected evaluation result.