ReplaySSM: cache inputs, rebuild state on the fly

As hybrid SSM-transformer models become common, avoiding per-step SSM state writes could make decode substantially faster without changing outputs

Why do we store the SSM state at all? More and more models are hybrids (Nemotron-3, Qwen3.5), so SSM decode speed matters. We only write it back every step so the next step can read it. ReplaySSM caches the recent inputs instead and rebuil