Struct KdaWeights 

pub struct KdaWeights {

    pub q_proj: QTensor,
    pub k_proj: QTensor,
    pub v_proj: QTensor,
    pub conv_q: Vec<f32>,
    pub conv_k: Vec<f32>,
    pub conv_v: Vec<f32>,
    pub f_a: QTensor,
    pub f_b: QTensor,
    pub dt_bias: Vec<f32>,
    pub a_log: Vec<f32>,
    pub b_proj: QTensor,
    pub gate: KdaOutGate,
    pub o_norm: Vec<f32>,
    pub o_proj: QTensor,
    pub gate_lower_bound: Option<f32>,
}

Fields

q_proj: QTensor

[nh·dk, hidden]

k_proj: QTensor

[nh·dk, hidden]

v_proj: QTensor

[nh·dv, hidden]

conv_q: Vec<f32>

[nh·dk × kk] — depthwise taps, oldest→newest (see GdnWeights.conv1d)

conv_k: Vec<f32>

conv_v: Vec<f32>

[nh·dv × kk]

f_a: QTensor

[rank, hidden] — low-rank decay projection, stage 1

f_b: QTensor

[nh·dk, rank] — stage 2

dt_bias: Vec<f32>

[nh·dk]

a_log: Vec<f32>

[nh] per-head (Kimi-Linear-48B) | [dk] per-dim (Kimi-K3) | [nh·dk] full — broadcast resolved by length.

b_proj: QTensor

[nh, hidden] — β = σ(b_proj·x) per head

gate: KdaOutGate

Output gate: full-rank g_proj (K3) or low-rank g_b(g_a(x)) (48B).

o_norm: Vec<f32>

[dv] — gated RMSNorm weight (per head over head_v_dim)

o_proj: QTensor

[hidden, nh·dv]

gate_lower_bound: Option<f32>

Some(lb): log-decay = lb·σ(exp(A)·(f+bias)) (K3, lb=−5); None: −exp(A)·softplus(f+bias) (Kimi-Linear-48B).