设计适合硬件运行的模型架构，并以最小的性能损失迁移现有LLM模型
Design hardware-friendly model architectures and migrate existing LLMs with minimal performance loss
