Infrastructure & deployment
Accelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tile
NVIDIA and Jane Street report cuTile forward-pass kernels for Mamba-2 that outperform a Triton reference in their tested configurations.
Why it matters: For engineers building sequence-model workloads, the post offers a concrete example of expressing chunked recurrent computation as tiled GPU kernels and tuning them with cuTile. The reported speedups are limited to forward-pass kernels, so they do not establish end-to-end training gains.