vLLM vs TensorRT-LLM vs TGI: how do they differ and when do you pick each?

The screen is whether you can name the real differentiators (continuous batching and PagedAttention in vLLM, compiled per-model engines in TensorRT-LLM, the operational packaging of TGI) rather than reciting throughput numbers off a vendor chart. Candidates who say 'TensorRT-LLM is fastest, use it' fail the follow-up about build time, model coverage, and hardware lock-in. The strong answer ties the pick to a concrete constraint: hardware, time-to-ship, or peak throughput on a fixed model.