transformers

Flash-attn performance: remove cuda sync during inference

#33570

Merged