Hi, I noticed the default per_device_train_batch_size in the training script is 1. When I increased it to 16 for better efficiency, the model performance (especially visual grounding accuracy) didn't ...