Struct QuantizedMatrixQ4 

pub struct QuantizedMatrixQ4 {
    pub data: Vec<u8>,
    pub scales: Vec<f32>,
    pub n: usize,
    pub k: usize,
}

A weight matrix quantized to symmetric int4, packed two values per byte.

Layout mirrors crate::int8::QuantizedMatrix: [n, k] row-major in the checkpoint’s own nn.Linear orientation, one f32 scale per output channel, so no transpose is ever materialized.

Fields

data: Vec<u8>

n * k.div_ceil(2) bytes: row-major, two biased nibbles per byte, low nibble first.

scales: Vec<f32>

One scale per output channel.

n: usize

k: usize

Implementations

impl QuantizedMatrixQ4

pub fn quantize(weight: &[f32], n: usize, k: usize) -> Self

Quantizes an [n, k] f32 weight matrix.

Panics

If weight.len() != n * k, or a weight is non-finite — a NaN reaching the quantizer means the graph upstream is already corrupt, and refusing loudly beats baking it into an artifact.

pub fn dequantize_row(&self, row: usize) -> Vec<f32>

Dequantizes one output channel back to f32, for parity comparison against the f32 weights.

pub fn packed_bytes(&self) -> usize

Bytes of weight storage, the number this lever exists to shrink.

Trait Implementations

impl Clone for QuantizedMatrixQ4

fn clone(&self) -> QuantizedMatrixQ4

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizedMatrixQ4

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

Formats the value using the given formatter.

impl PartialEq for QuantizedMatrixQ4

fn eq(&self, other: &QuantizedMatrixQ4) -> bool

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for QuantizedMatrixQ4