Classifier free guidance mask

[sanagno]

The caption embedder during training randomly drops some captions here to enable classifier free guidance. Shouldn't the mask be updated in such a case so that all tokens (from the y_embedder.y_embedding) at attended at?

Same should hold during inference, where instead of repeating the mask as done here, a mask of all 1's should be interleaved.

The current implementation is though consistent between training and inference.