transformers

Fix `torch.compile` with `fullgraph=True` when `attention_mask` input is used

#29211

Merged