LLM Inference Speed for Text Generation