Struct PerLayerQuant 

pub struct PerLayerQuant {
    pub trits: Vec<Trit>,
    pub scale: f32,
    pub shape: Vec<usize>,
    pub name: String,
    pub sparsity: f64,
}

The result of quantizing one weight tensor.

Fields

trits: Vec<Trit>

Ternary weights in row-major order.

scale: f32

Per-layer scale factor α = mean(|W|). Multiply by this to approximate the original magnitudes.

shape: Vec<usize>

Shape: [rows, cols] for 2D tensors; arbitrary for higher ranks.

name: String

Human-readable layer name (e.g. “model.layers.0.self_attn.q_proj.weight”).

sparsity: f64

Fraction of trits that are Tend (zero) — 0.0 … 1.0.

Implementations

impl PerLayerQuant

pub fn quantize( weights: &[f32], name: impl Into<String>, shape: Vec<usize>, ) -> Self

Quantize weights (flat, row-major) to balanced ternary.

name — layer identifier for diagnostics. shape — logical shape of the tensor (product must equal weights.len()).

pub fn reconstruct(&self) -> Vec<f32>

Approximate reconstruction: W_approx = α × W_t (as f32). Useful for measuring quantization error.

pub fn mse(&self, original: &[f32]) -> f32

Mean squared error between original weights and ternary reconstruction.

Trait Implementations

impl Clone for PerLayerQuant

fn clone(&self) -> PerLayerQuant

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for PerLayerQuant

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

Formats the value using the given formatter.