Struct ArchSpec 

pub struct ArchSpec {
    pub activation: Activation,
    pub norm_flavor: NormFlavor,
    pub qk_norm: bool,
    pub sandwich_norms: bool,
    pub embed_scale_sqrt_hidden: bool,
    pub attn_logit_softcap: Option<f64>,
    pub final_logit_softcap: Option<f64>,
    pub query_pre_attn_scalar: Option<f64>,
    pub rope_theta: f64,
    pub rope_scaling: RopeScaling,
    pub rope_local_theta: Option<f64>,
    pub layers: Vec<LayerKind>,
}

Resolved architecture description consumed by the decoder.

Fields

activation: Activation

norm_flavor: NormFlavor

qk_norm: bool

Per-head RMSNorm on Q and K after projection (gemma-3, qwen-3).

sandwich_norms: bool

Extra post-attention / post-MLP norms around the residual adds (gemma-3 “sandwich” norms).

embed_scale_sqrt_hidden: bool

Multiply embeddings by sqrt(hidden_size) (gemma; computed in f32).

attn_logit_softcap: Option<f64>

Soft-capping of attention logits / final logits (gemma-2 family).

final_logit_softcap: Option<f64>

query_pre_attn_scalar: Option<f64>

Attention scale divisor override (1/sqrt(query_pre_attn_scalar) instead of 1/sqrt(head_dim)).

rope_theta: f64

Global-layer RoPE base + scaling.

rope_scaling: RopeScaling

rope_local_theta: Option<f64>

Sliding layers rotate with this base instead (gemma dual-RoPE).

layers: Vec<LayerKind>

One entry per transformer layer.

Implementations

impl ArchSpec

pub fn resolve(meta: &ModelMetadata) -> Self

pub fn windows(&self) -> Vec<Option<usize>>

Per-layer window sizes in the shape PagedKVCache::new_with_windows takes (None = global arena, Some(w) = rolling window).

Trait Implementations

impl Clone for ArchSpec

fn clone(&self) -> ArchSpec

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ArchSpec

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

Formats the value using the given formatter.