passing attention mask to PPOTrainer

[HassanJbara]

Is there a way to pass attention masks of tokenized queries to the PPOTrainer or the generate() function? I see that other classes in the library have padding data collators, where attention masks could be included in the dataset, but not this one.