TensorRT-LLM
The last few percent of GPU throughput, paid for in build complexity and hardware lock-in.
Where vLLM and SGLang give you a strong default quickly, TensorRT-LLM asks you to compile an engine per model, per GPU configuration, per precision — and gives you throughput at the top of what the hardware can do. That trade makes sense when you are serving one or two models at very high volume on known hardware. It makes much less sense during evaluation, when model and hardware are still changing weekly. Expect the build step to be part of your deployment pipeline, not a one-off.
Best for: High-volume serving of a stable model set on fixed NVIDIA hardware