11,840 TPS on a single GPU

Nearly doubling Nemotron throughput on one GB300, with interactive explainers for speculative decoding, scheduling, and self-tuning inference.