Build a unified pipeline that optimizes LLMs using both structured pruning and dynamic prompt compression via compressed sensing. Developers need a way to chain these two techniques without destroying model performance during inference.