Large language models (LLM) such as LLaMA 2 and Falcon can require dozens, if not hundreds, of gigabytes of GPU memory.