Generative AI workloads place unusual demands on system architecture: inference is stateful, memory-bandwidth-bound, and bursty, and a single large model can saturate expensive accelerators while still leaving them underutilized. This article catalogs the solution-architecture patterns that have converged in production large-language-model (LLM) systems and organizes them into a layered reference architecture spanning the model gateway, orchestration, knowledge and caching, and the inference-serving cluster. We treat the core scaling primitives — continuous (in-flight) batching, PagedAttention-style KV-cache management, prefix and prompt caching, tensor parallelism, quantization, and speculative decoding — as composable patterns, and give the quantitative relationships that govern their cost and performance, including KV-cache memory growth, the throughput–latency trade-off, and an arithmetic-intensity (roofline) view of the prefill versus decode phases. We also cover application-level patterns (retrieval-augmented generation, agent orchestration, the model gateway, and semantic caching) and operational concerns (autoscaling, multi-tenancy, and reliability). The aim is a practitioner's map: which pattern solves which bottleneck, what it costs, and how the patterns combine. Public technical sources are cited throughout.
Sushma Sunkollu Nagaraj (Tue,) studied this question.