pub fn launch_k3_kda_decode_token(
gpu: &dyn GpuBackend,
kernels: &K3KdaDecodeKernels,
x_qkv: &[f32],
conv_w: &[f32],
gate: &[f32],
beta: &[f32],
cfg: &KdaConfig,
state: &mut KdaState,
stream: u64,
) -> Result<Vec<f32>>Expand description
One-token CUDA KDA decode. Host-state round-trip (oracle / CPU escape).