LoRA (Low-Rank Adapter) is way to customize/finetune the LLM to a new datasets without needing to retrain the entire network (which makes it better (and in theory easier to do). It doesn't not change the speed significantly afaik