Struct KdaAttnLayer 

pub struct KdaAttnLayer {

    pub plan: KimiDeltaNetPlan,
    pub wq: GpuTensor,
    pub wk: GpuTensor,
    pub wv: GpuTensor,
    pub f_a: GpuTensor,
    pub f_b: GpuTensor,
    pub g_a: GpuTensor,
    pub g_b: GpuTensor,
    pub b_proj: GpuTensor,
    pub wo: GpuTensor,
    pub conv: CudaSlice<f32>,
    pub a_log: GpuTensor,
    pub dt_bias: GpuTensor,
    pub o_norm: GpuTensor,
    pub tp: Option<Box<Glm5TpKda>>,
}

One loaded KDA mixer. Field names follow the reference’s tensor roles, not the HF spellings.

Fields

plan: KimiDeltaNetPlan

wq: GpuTensor

q/k/v projections, [qkv, hidden] each.

wk: GpuTensor

wv: GpuTensor

f_a: GpuTensor

Forget gate low-rank pair: f_a [head_dim, hidden], f_b [qkv, head_dim].

f_b: GpuTensor

g_a: GpuTensor

Output gate low-rank pair, same shapes as the forget pair.

g_b: GpuTensor

b_proj: GpuTensor

Per-head beta projection, [heads, hidden].

wo: GpuTensor

Output projection, [hidden, qkv].

conv: CudaSlice<f32>

The three per-plane conv weights concatenated into [3*qkv, kernel] (see module header).

a_log: GpuTensor

A_log [heads], dt_bias [qkv] (per CHANNEL, unlike GDN’s per-head bias), o_norm [head_dim].

dt_bias: GpuTensor

o_norm: GpuTensor

tp: Option<Box<Glm5TpKda>>

glm5 TP-2 sidecar (MEMRA_GLM5_TP, lane/glm5-tp2). Some means THIS layer struct is the ROOT-RANK HEAD SHARD (heads/2) and the sidecar carries the peer shard + runtime. Every plain entry point REFUSES a sharded layer by name — only the TP walk (glm5_tp::kda_tp_*) may execute it. None everywhere else (zero cost, zero change).

Implementations

impl KdaAttnLayer

pub fn heads(&self) -> usize

pub fn head_dim(&self) -> usize

pub fn qkv(&self) -> usize

pub fn conv_kernel(&self) -> usize

pub fn conv_width(&self) -> usize

Fused conv ring width, matching StatePlan::Recurrent { conv_width } for this layer.

pub fn state_width(&self) -> usize

Recurrent state elements, matching StatePlan::Recurrent { state_width }.

pub fn load( e: &Engine, src: &dyn TensorSource, il: u32, plan: &KimiDeltaNetPlan, ) -> Result<Self, Box<dyn Error>>

Load block il’s KDA tensors. Names are the ggml-dialect contract names from memra_gguf::tensor_contract::add_kda; the safetensors source translates them.