ReplaySSM accelerates hybrid model decoding by caching inputs instead of state
NVIDIA and Princeton researchers introduce ReplaySSM, a technique that speeds up State Space Model decoding by reducing memory traffic and simplifying speculative rollback.
