Not quite a drop in replacement, but close enough.
From the paper[1]:
> Llama 2, an updated version of Llama 1, trained on a new mix of publicly available data. We also increased the size of the pretraining corpus by 40%, doubled the context length of the model, and adopted grouped-query attention (Ainslie et al., 2023)[2].
[1]: https://ai.meta.com/research/publications/llama-2-open-found...
[2]: https://arxiv.org/abs/2305.13245
