Struct K3LayerCtx 

pub struct K3LayerCtx<'a> {
    pub kda: &'a KdaConfig,
    pub mla: &'a MlaConfig,
    pub moe: &'a LatentMoeConfig,
    pub situ_beta: f32,
    pub situ_linear_beta: f32,
    pub hidden: usize,
    pub dense_intermediate: usize,
    pub eps: f32,
    pub rope_theta: f32,
    pub reduce_hidden: Option<HiddenReduce<'a>>,
    pub dense_mlp: Option<DenseMlpCore<'a>>,
}

Per-layer geometry the GPU wrapper and forward_token share.

K3BoundLayer::decode copies hidden D2H, runs this math, copies H2D. LinearAttention injects CUDA conv+recurrent unless K3_CUDA_KDA=0. FullAttention injects CUDA gated-NoPE MLA unless K3_CUDA_MLA=0.

Fields

kda: &'a KdaConfig

mla: &'a MlaConfig

moe: &'a LatentMoeConfig

situ_beta: f32

situ_linear_beta: f32

hidden: usize

dense_intermediate: usize

eps: f32

rope_theta: f32

reduce_hidden: Option<HiddenReduce<'a>>

After row-parallel o_proj (and dense MLP down). None at TP=1.

dense_mlp: Option<DenseMlpCore<'a>>

Implementations

impl<'a> K3LayerCtx<'a>

pub fn from_model(model: &'a K3CpuModel) -> Self