This package initializer exposes three named implementations through kimi_k3.attention: KDARecurrent, KDAChunkwise, and GatedMLA. Its full implementation consists of three relative imports source L1–L3.
Public import surface
from .kda.recurrent import KDARecurrent
from .kda.chunkwise import KDAChunkwise
from .gated_mla import GatedMLAA from kimi_k3.attention import KDAChunkwise resolves the reference KDA implementation from attention/kda/chunkwise.py, while KDARecurrent provides token-by-token reference behavior. GatedMLA is implemented in attention/gated_mla.py. The imports are eager: loading this subpackage loads those Python modules.
What is not exported here
FusedKDAChunkwise is not a package-level export. The model factory imports it lazily when kda_implementation == "fused" test test_kimi_k3.py L57–L62; see model.py and attention/kda/fused_chunk.py for the selection and CUDA/FLA dependency. No __all__ is declared here, and no attention computation or state management takes place in the initializer.