ReplaySSM cuts state writes, not state reads
NVIDIA and Princeton report faster hybrid-model decoding by checkpointing recurrent state and replaying recent inputs. The deployment decision turns on batch size, rollback, and precision, not the headline speedup.