transformers

Fix llama model sdpa attention forward function masking bug when output_attentions=True

#30652

Merged