Struct V3Config

Source

pub struct V3Config {Show 16 fields
    pub vocab_size: usize,
    pub out_vocab: usize,
    pub d_model: usize,
    pub num_heads: usize,
    pub num_kv_heads: usize,
    pub num_layers: usize,
    pub qk_head_dim: usize,
    pub v_head_dim: usize,
    pub max_seq_len: usize,
    pub hada_n: usize,
    pub mhc_lanes: usize,
    pub rope_theta: f32,
    pub sliding_window: usize,
    pub global_layers: Vec<usize>,
    pub qkv_conv_taps: usize,
    pub engram: V3Engram,
}

Expand description

Full v3 model geometry.

Fields§

§vocab_size: usize§out_vocab: usize

Rows of the tied head actually used for logits. Rows beyond this are input-only code embeddings. 0 means the full vocabulary.

§d_model: usize§num_heads: usize§num_kv_heads: usize§num_layers: usize§qk_head_dim: usize

Query and key head width — 48 on the shipped model.

§v_head_dim: usize

Value head width — 64. Unlike v2, this differs from qk_head_dim.

§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§sliding_window: usize

Local attention width for layers that are not global. 0 means full causal everywhere.

§global_layers: Vec<usize>

Layers attending over the whole sequence instead of sliding_window.

§qkv_conv_taps: usize

Causal depthwise convolution width over Q, K and V. 0 means none.

§engram: V3Engram

Implementations§

Source §

impl V3Config

Source

pub fn q_dim(&self) -> usize

Width of the concatenated query projection, num_heads * qk_head_dim.

Source

pub fn k_dim(&self) -> usize

Width of the concatenated key projection.

Source

pub fn v_dim(&self) -> usize

Width of the concatenated value projection.

Source

pub fn attn_out_dim(&self) -> usize

Width of the attention output before out_proj, num_heads * v_head_dim.

Source

pub fn kv_repeat(&self) -> usize

Query heads served by one key/value head.

Source

pub fn mhc_width(&self) -> usize

Flattened mHC lane-stream width, mhc_lanes * d_model.

Source

pub fn logit_rows(&self) -> usize

Rows of the tied head used for logits.

Source

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

True when layer attends over the whole sequence.

Source

pub fn attention_span(&self, layer: usize) -> Option<usize>

How many past positions layer can attend to, counting the current one. None means unbounded — the caller must keep full history.

This is the distinction that drives the KV allocation: local layers need a fixed ring, global layers need everything.

Source

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

The lane layer i is assigned to, as eye(n)[i % n] in the reference Stack.

Source

pub fn kv_bytes(&self, seq_len: usize, bytes_per: usize) -> usize

Bytes of key/value cache for a session of seq_len positions, at bytes_per per stored scalar.

Local layers cost a fixed ring; global layers cost the real sequence length, which is why they are sized to what a session actually uses rather than to max_seq_len.

Source

pub fn kv_bytes_at(&self, seq_len: usize, precision: KvPrecision) -> usize

Bytes of key/value cache for a session of seq_len positions at a given storage precision.

Prefer this to Self::kv_bytes when the answer decides whether a session fits: the int8 figure is not simply a quarter, because each stored head vector also carries an f32 scale.