What it is
FlashKDA implements the forward pass of Kimi Delta Attention, a linear attention variant, as CUTLASS-based CUDA kernels for SM90 and newer GPUs. It supports bf16 inputs, recurrent initial and final states, and variable-length batching through cumulative sequence lengths. Once installed it is picked up automatically by flash-linear-attention's chunk_kda during inference, so model developers and inference engineers serving Kimi-style linear attention models get faster prefill without code changes.