Llama 4: eager attention results in wrong casual mask shape

[Qubitium]

System Info

Latest transformers 4.51.0

Who can help?

@ArthurZucker @SunMarc @MekkCyber I know you guys are probably burning an all-nighter due to Llama 4. =)

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

While adding GPTQModel support Llama 4 which does transformers inference at the layer/module level, we found issues where when eager attention is used, casual mask shapes are off by exactly factor of 2 or 4 for Llama 4:

ModelCloud/GPTQModel#1508

We are able to bypass this for now by setting batch=1, removing padding, and not passing attention_mask for inference.

Expected behavior

No error.

[ArthurZucker]

Thanks will have a look asap!