Q: Padded tokens without an `attention_mask`?

[Axe--]

Hey, I wanted to know how the padded tokens are handled by the attention layers of the model,
considering that repo's implementation of GPT-2 discards attention_mask (HuggingFace).