Module kda_cuda 

Host launch for K3 CUDA KDA decode (kda_decode PTX module).

Two kernels, one token: conv-4 + SiLU, then L2 q/k + delta-rule. CPU oracle: atlas_core::kimi_k3::kda_decode_token. BoundLayer serve LinearAttention default is this launch (K3_CUDA_KDA=0 keeps CPU).

Structs

K3KdaDecodeKernels

KdaDeviceState

Device-resident conv + recurrent. Seed once; decode does not D2H them.

Constants

CONV_ENTRY

MODULE

PTX module stem = kernels/gb10/kimi-k3/bf16/kda_decode.cu.

RECURRENT_ENTRY

Functions

launch_k3_kda_decode_token

One-token CUDA KDA decode. Host-state round-trip (oracle / CPU escape).

launch_k3_kda_decode_token_on_device

One-token CUDA KDA on device-resident conv/recurrent. Does not D2H state.