Struct V2Config 

pub struct V2Config {
    pub vocab_size: usize,
    pub d_model: usize,
    pub num_heads: usize,
    pub num_kv_heads: usize,
    pub num_layers: usize,
    pub head_dim: usize,
    pub max_seq_len: usize,
    pub hada_n: usize,
    pub mhc_lanes: usize,
    pub rope_theta: f32,
    pub kv_window: usize,
    pub engram: EngramGeometry,
}

Full v2 model geometry.

Fields

vocab_size: usize

d_model: usize

num_heads: usize

num_kv_heads: usize

num_layers: usize

head_dim: usize

max_seq_len: usize

hada_n: usize

Hadamard width in HadamardMLP: d_model rounded up to a power of two.

mhc_lanes: usize

rope_theta: f32

kv_window: usize

Sliding attention window. 0 disables it (full causal).

engram: EngramGeometry

Implementations

impl V2Config

pub fn attn_dim(&self) -> usize

Attention width, num_heads * head_dim.

pub fn kv_dim(&self) -> usize

Packed K/V projection width.

pub fn kv_repeat(&self) -> usize

Query heads served by one KV head.

pub fn mhc_width(&self) -> usize

Flattened mHC lane-stream width, mhc_lanes * d_model.

pub fn active_lane(&self, layer: usize) -> usize

The lane layer i is “assigned” to — eye(n)[i % n] in architecture.Stack, which drives the pre_off / post_off biases.

pub fn pre_off(&self, layer: usize, lane: usize) -> f32

pre_off[layer][lane] — 8 * lane_onehot - 4.

pub fn post_off(&self, layer: usize, lane: usize) -> f32

post_off[layer][lane] — -4 * (1 - lane_onehot).

pub fn validate(&self) -> Result<(), &'static str>

Trait Implementations

impl Clone for V2Config

fn clone(&self) -> V2Config

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for V2Config

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

Formats the value using the given formatter.

impl PartialEq for V2Config

fn eq(&self, other: &V2Config) -> bool

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for V2Config