EOS stopping issue

[yxs87]

Hello.There seems to be an inconsistency between training and inference regarding the EOS stopping mechanism. While the model is trained with EOS tokens to indicate the end of a sequence, this mechanism isn't applied during the predict_test phase. As a result, the model continues predicting up to the maximum number of steps (e.g., over 200 tokens), which unnecessarily prolongs the inference time for each image.