Dynamic Activation Pitfalls in LLaMA Models: An Empirical Study