FlashKDA open-sourced: high-performance CUTLASS-based attention kernels with 1.72-2.22x prefill speedup

FlashKDA is a CUTLASS-based implementation of Kimi Delta Attention, delivering 1.72-2.22x prefill speedup on H20 and works as a drop-in backend for flash-linear-attention.

• open-source
• FlashKDA
• attention kernel
• Kimi Delta Attention