Best practice for training LLaMA models in Megatron-LM - View it on GitHub