Engines compared from the inside: llama.cpp, vLLM, SGLang, TGI, TensorRT-LLM, Dynamo

How the major inference engines differ in architecture rather than marketing: scheduling and KV management, kernels, quantisation formats, hardware targets, structured decoding, multi-GPU strategy, and a decision framework with the questions that matter.