1. phase1 and phase2 are finetuned based on Qwen 32B VL
2. add phase1 and phase2 model in inference.sh
3. vllm start: nohup bash service.sh 4 > vllm.log 2>&1 &
4. bash inference.sh
5. phase1 and phase2 models are on baidu cloud: https://pan.baidu.com/s/1Zr7IFDTdNhePbS1LCsH3Yw?pwd=EuWX, 提取码：EuWX The model should be merged with offical Qwen2.5-VL-32B-Instruct which please be downloaded yourself, because it is very large. After merging, the model would be used for evaluation, and merge.py file is put on the above baidu cloud link.