For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo module

kda

Graph-op bindings for the KDA (Kimi Delta Attention) kernels.

The kernel math and kda_chunk's host-side launch sequence live in //Kernels/lib/kda (kda.recurrent, kda.chunk_fwd, kda.chunk_launch); only the @extensibility.register wrappers live here, mirroring the msa.mojo binding for //Kernels/lib/msa. Registration has to be declared inside the built-in kernel library itself -- importing the struct from the standalone lib does not put the op in the graph compiler's registry, so a served graph resolving kda_decode / kda_chunk needs this declaration and not just the dependency.

Two ops, one per prefill regime:

• kda_decode -- the token-sequential recurrence (kda.recurrent). Correct for any ragged batch (decode's 1-token sequences and a multi-token prefill alike), but O(T) sequential steps per sequence.
• kda_chunk -- the chunk-parallel prefill pipeline (kda.chunk_fwd, driven by kda.chunk_launch): L1 kda_chunk_prepare_gpu (parallel per chunk) -> L2 kda_chunk_scan_gpu (sequential, but only O(T/CHUNK_SIZE) steps) -> L3 kda_chunk_output_gpu (parallel). This is what turns prefill from O(T) into O(T/CHUNK_SIZE) sequential depth; the Python KDA layer picks it for prefill and keeps kda_decode for single-token decode (see kda_attention.py).

Structs

• ​KdaChunk: KDA chunk-parallel prefill forward pass.
• ​KdaDecode: KDA decode recurrence forward pass.