From Large to Super-Tiny: End-to-End Optimization for Cost-Efficient LLMs