ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization