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. (Read More)Source