Struct MlaDims 

pub struct MlaDims {
    pub n_head: usize,
    pub d_nope: usize,
    pub d_rope: usize,
    pub d_v: usize,
    pub kv_rank: usize,
}

MLA head geometry. GLM-5.2: n_head=64, d_nope=192, d_rope=64, d_v=256, kv_rank=512.

Fields

n_head: usize

d_nope: usize

qk nope head dim (P)

d_rope: usize

qk rope head dim (R); latent cache row = kv_rank + d_rope

d_v: usize

v head dim (V)

kv_rank: usize

kv lora rank (Lkv)

Implementations

impl MlaDims

pub const GLM52: MlaDims

pub const GLM5_NEXT: MlaDims

glm5_next (GLM-5.3-Flash) MLA geometry: NoPE — rope_head_dim is 0, so there is no decoupled rope plane at all and the latent cache row is kv_rank wide (not kv_rank+rope). qk_head_dim is therefore ALL nope (256), and the softmax scale is 1/sqrt(256) = 1/16 — numerically the same 1/16 as GLM-5.2 (192+64), reached by a different decomposition.

pub fn scale(&self) -> f32

Softmax scale: 1/sqrt(d_nope + d_rope) — the ORIGINAL qk head dim (256 for GLM-5.2), NOT the absorbed width (576). llama.cpp glm-dsa.cpp kq_scale with mscale=1 (no yarn).

Trait Implementations

impl Clone for MlaDims

fn clone(&self) -> MlaDims

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for MlaDims

impl Debug for MlaDims

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

Formats the value using the given formatter.