DeepSpeed LLM An example how to train LLMs efficiently on a small GPU using DeepSpeed Installation conda create -n deepspeed python=3.8