SpecEE: Accelerating Large Language Model Inference with Speculative Early Exiting