Struct Mla 

pub struct Mla {
    pub heads: usize,
    pub q_lora_rank: Option<usize>,
    pub kv_lora_rank: usize,
    pub qk_nope: usize,
    pub qk_rope: usize,
    pub v_dim: usize,
    pub rope: Rope,
    pub scale: f64,
    pub kv_b_split: bool,
}

Multi-head latent attention (DeepSeek V2/V3): the keys and values of every head are expanded from one shared low-rank latent per position (kv_lora_rank wide, RMS-normed), and each head’s query and key carry a further qk_rope dimensions that rotate — the key’s shared by every head. The query is a full projection, or itself low-rank (q_lora_rank, normed between its two halves).

Fields

heads: usize

q_lora_rank: Option<usize>

kv_lora_rank: usize

qk_nope: usize

Per-head query/key dimensions that do not rotate.

qk_rope: usize

Per-head query/key dimensions that rotate (the trailing ones).

v_dim: usize

rope: Rope

scale: f64

The score scale: 1/√(qk_nope + qk_rope), times YaRN’s magnitude squared.

kv_b_split: bool

The checkpoint stores the latent’s expansion split into its key half (attn_k_b, each head’s kv_lora_rank rows of qk_nope) and its value half (attn_v_b, each head’s v_dim rows of kv_lora_rank), rather than as one attn_kv_b of qk_nope + v_dim rows a head.

Trait Implementations

impl Clone for Mla

fn clone(&self) -> Mla

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Mla

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

Formats the value using the given formatter.

impl PartialEq for Mla

fn eq(&self, other: &Mla) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for Mla