Struct Glm5NextKdaConfig 

pub struct Glm5NextKdaConfig {
    pub hidden: usize,
    pub heads: usize,
    pub head_dim: usize,
    pub conv_kernel: usize,
    pub gate_lower_bound: f32,
    pub rms_norm_eps: f32,
    pub l2_eps: f32,
    pub chunk: usize,
}

2026-09-25: KDA geometry and the two numeric config values the block uses. The loader fills every field from ModelConfig except l2_eps and chunk; there is no Default.

Fields

hidden: usize

heads: usize

head_dim: usize

conv_kernel: usize

gate_lower_bound: f32

2026-09-25: linear_attn_config.gate_lower_bound.

rms_norm_eps: f32

2026-09-25: rms_norm_eps, read by o_norm.

l2_eps: f32

2026-09-25: The L2 epsilon, applied as 1/sqrt(sum + eps) rather than max(norm, eps). Not a config key.

chunk: usize

2026-09-25: Prefill chunk width, bounded by SMEM_CEILING through validate.

Implementations

impl Glm5NextKdaConfig

pub fn qkv_dim(&self) -> usize

pub fn conv_dim(&self) -> usize

2026-09-25: q | k | v concatenated: the channel count the depthwise conv sees.

pub fn qk_channels(&self) -> usize

2026-09-25: The q | k channels, the only ones L2-normalised.

pub fn recurrent_state_elems(&self) -> usize

2026-09-25: Elements of the FP32 [heads, head_dim, head_dim] K-major recurrent state.

pub fn conv_state_elems(&self) -> usize

2026-09-25: Elements of the FP32 [conv_dim, conv_kernel] conv state (see KdaSeqState).

pub fn smem_prepare(&self) -> usize

pub fn smem_scan(&self) -> usize

pub fn validate(&self) -> Result<()>

Trait Implementations

impl Clone for Glm5NextKdaConfig

fn clone(&self) -> Glm5NextKdaConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Glm5NextKdaConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Copy for Glm5NextKdaConfig