FlashKDA
High-performance Kimi Delta Attention kernels
...The package integrates with flash-linear-attention and can be selected automatically as the backend for chunk_kda during inference. It supports recurrent state input and output, variable-length batches, internal gating, query-key normalization, and beta activation. Builds can target the detected GPU architecture or multiple supported architectures for wheels and CI pipelines. The repository also provides correctness tests, benchmark material, a direct Python kernel API, and development helpers for CUDA and C++ tooling.