LUT-LLM: Efficient Large Language Model Inference with Memory-based Computations on FPGAs