Missing attention mask

[LaurentMazare]

Backend impacted

The MLX implementation

Operating system

Linux

Hardware

Metal with MLX

Description

There is an issue with the current mlx transformer implementation, the attention mask used in the transformer here turns out never to be set.
This is not an issue with the current inference set up as only the big transformer and the depformer use this implementation (and not the codec) and in these cases data is processed one step at a time so there is no need for any mask.

Extra information

.

Environment

.