I looked at their paper and the biggest gains come when you have many questions from one context, which is the way these models are supposed to be used. It's around 1.5-3x faster than vLLM with single questions. That's the worst case where you can't reuse anything.

The way I understand it most of their speedup comes from caching intermediate results. They freeze the state after processing context then branch each question off that point. It means they can process questions in parallel after the context stage is finished, and don't waste time doing the same work N times.