The paper introduces a formal resource‑allocation framework for serving large transformer‑based models with pipeline parallelism, proposing greedy placement, cache allocation, and load‑balancing algorithms that drastically cut inference latency.