Struct Glm5NextKdaLayer 

pub struct Glm5NextKdaLayer {
    pub layer_idx: usize,
    pub cfg: Glm5NextKdaConfig,
    pub weights: Glm5NextKdaWeights,
    pub kernels: Glm5NextKdaKernels,
}

2026-09-25: One bound KDA attention block.

Fields

layer_idx: usize

2026-09-25: The checkpoint layer index this block was bound from.

cfg: Glm5NextKdaConfig

weights: Glm5NextKdaWeights

kernels: Glm5NextKdaKernels

Implementations

impl Glm5NextKdaLayer

pub fn decode( &self, gpu: &dyn GpuBackend, hidden: DevicePtr, state: &KdaSeqState, ws: &Glm5NextKdaWorkspace, stream: u64, ) -> Result<()>

2026-09-25: Single-token decode, carrying both states. The result lands in ws.final_out; state is updated in place.

pub fn decode_k( &self, gpu: &dyn GpuBackend, hidden: DevicePtr, k: usize, state: &KdaSeqState, ws: &Glm5NextKdaWorkspace, snapshots: &[(DevicePtr, DevicePtr)], stream: u64, ) -> Result<()>

2026-09-25: k tokens of one sequence: the projections batched, the recurrence one row at a time. Used for the speculative verify, and for prefill sub-chunks that do not take the chunked arm (glm5next_layer/steps/forward.rs).

front_end and back_end run once over all k rows. For 2 <= k <= DENSE_GEMV_BATCHM_MAX_M on a target with dense_gemv_bf16_batchm, the output matches k serial Self::decode calls bit for bit: the batched GEMV gives each row the bits of the M = 1 GEMV, the pack, gate, sigmoid and o_norm kernels treat rows independently, and stateful_row walks the state one row at a time as decode does.

snapshots[t], (h_dst, conv_dst), receives the state after row t; rows past snapshots.len() take no snapshot. Refuses k == 0 or k above the workspace.

impl Glm5NextKdaLayer

pub fn prefill( &self, gpu: &dyn GpuBackend, hidden: DevicePtr, t: usize, state: &KdaSeqState, ws: &Glm5NextKdaWorkspace, stream: u64, ) -> Result<()>

2026-09-25: Chunked prefill over t tokens, starting from the carried state.

pub fn prefill_with_pad_fill( &self, gpu: &dyn GpuBackend, hidden: DevicePtr, t: usize, state: &KdaSeqState, ws: &Glm5NextKdaWorkspace, pad_fill: f32, stream: u64, ) -> Result<()>

2026-09-25: Self::prefill with the padded q/k/v tails filled with pad_fill.

Self::prefill passes zero. The layer harness passes other values to show that kda_chunk_* ignores positions at or past t in-kernel, which a zeroed tail cannot show.

impl Glm5NextKdaLayer

pub fn new( layer_idx: usize, cfg: Glm5NextKdaConfig, weights: Glm5NextKdaWeights, kernels: Glm5NextKdaKernels, ) -> Result<Self>