Unexpected bug for generate function

[qiuzhen8484]

Hello, when I use this model to tackle VQA task by passing the visual querys to inputs_embeds argument without input_ids, it results in that there is an extra dimension in attention_mask and position_ids, e.g., 442 vs 441 for last dimension. How can I fix it? Thanks.