#cuda

FlashKDA is a fast CUDA kernel library for Kimi Delta Attention, built with CUTLASS and made for SM90+ GPUs with CUDA 12.9+ and PyTorch 2.4+. It can plug into flash-linear-attention as a backend, and on H20 it reports about 1.72× to 2.22× faster prefill than the existing baseline. This can help you run attention faster, with less deployment work, especially if you use supported hardware and the fixed K = V = 128 setup.

https://github.com/MoonshotAI/FlashKDA