transformers

628c54c2 - perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models (#45555)

perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models (#45555) * perf: avoid redundant rotary_emb calls for `position_embeddings` * perf: avoid redundant rotary_emb calls for `position_embeddings` dict