GPU Kernels
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.