Mamba 2
NVIDIA details cuTile kernels for Mamba-2 forward pass
NVIDIA outlined three cuTile kernels for Mamba-2 and said optimized forward-pass code ran 1.14x to 1.82x faster than a Triton reference.
Mamba 2
NVIDIA and Princeton Introduce ReplaySSM to Accelerate Hybrid AI Model Decoding
ReplaySSM addresses state-space model bottlenecks by caching inputs instead of state, delivering up to 1.8× faster decoding and 2.1× speculative serving throughput on production models.