Module mla_cuda 

Host launch for K3 CUDA gated-NoPE MLA decode (mla_decode PTX module).

Two kernels, one token: maybe_rope (NoPE skips rotate), then SDPA + gate. CPU oracle: atlas_core::kimi_k3::mla_decode_token. BoundLayer serve FullAttention default is this launch (K3_CUDA_MLA=0 keeps CPU).

Structs

K3MlaDecodeKernels

MlaDeviceKv

Device-resident MLA KV. Append is one-row D2D/H2D; SDPA reads the buffer. Host MlaKv remains the CPU oracle. Do not re-upload [0..T] each token.

Constants

MODULE

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

ROPE_ENTRY

SDPA_ENTRY

Functions

launch_k3_mla_decode_token

One-token CUDA gated-NoPE MLA. Updates q/k (rope) and kv (append).

launch_k3_mla_decode_token_on_device

Device-resident KV: rope stays on device, append is one row, one D2H (output).