Transformer-Lite: High-efficiency Deployment of Large Language Models on Mobile Phone GPUs