MoonshotAI Launches FlashKDA: Open-Source High-Performance Attention Kernel Now Live on GitHub

MoonshotAI launches FlashKDA, an open-source high-performance attention kernel built on CUTLASS, now live on GitHub with 819 stars, supporting CUDA SM90+ architectures and seamlessly integrating with the flash-linear-attention library for cutting-edge AI inference.

• MoonshotAI releases FlashKDA, a high-performance open-source implementation of Kimi Delta Attention kernels built on CUTLASS, now available on GitHub with 819 stars and 77 forks.
• FlashKDA supports CUDA architectures SM90 and above, integrates seamlessly as an auto-dispatched backend for the flash-linear-attention library, and handles variable-length batching through cumulative sequence length inputs.
• The kernel exposes a forward pass API accepting queries, keys, values, gates, and optional recurrent states, currently requiring head dimensions K = V = 128, with benchmarks available for H20 and GB200 hardware.