Large language model (LLM) inference is shifting from a local engine optimization problem to a distributed control problem that involves reusable state, phase placement, heterogeneous accelerators, networking, autoscaling, reliability and service‑level objectives. This article maps that transition across peer‑reviewed systems research, open‑source projects and documented production studies, treating vLLM and llm‑d as complementary layers.\ \ Model‑serving engines improve execution through mechanisms such as PagedAttention, continuous batching, specialized kernels, quantization and parallelism, while an inference control plane decides where, when and under which policy the execution occurs across a fleet, covering aggregated versus disaggregated topology, KV source and transfer, hardware variant, routing/admission policy and slower scaling decisions.\ \ The contribution is a synthesis rather than a new benchmark; all reported performance and deployment numbers remain attributed to their original sources. The combined evidence indicates that the scarce resource in modern inference is moving from raw FLOPs toward managed state, placement, network movement, reliability and decision quality.\ \ We propose an Inference Execution Planner that selects feasible execution plans instead of only endpoints, taking into account the aforementioned dimensions. The paper also supplies a source‑level benchmark atlas, a bottleneck‑migration taxonomy, practical deployment guidance, an evaluation framework based on SLO‑goodput, and research questions for agentic, multimodal, heterogeneous and resilient inference.\ \ Review