Struct Int4EmbeddingsView 

pub struct Int4EmbeddingsView<'a> {
    pub scales: &'a [u8],
    pub codes: &'a [u8],
    pub n: usize,
    pub dim: usize,
    pub blocks: usize,
}

Decoded view over an int4 embeddings payload. Slices borrow the input bytes (no copy); accessors decode scales/codes on demand.

Fields

scales: &'a [u8]

f16 LE group scales, n * blocks of them, row-major.

codes: &'a [u8]

packed nibble codes, n * dim/2 bytes.

n: usize

dim: usize

blocks: usize

Implementations

impl<'a> Int4EmbeddingsView<'a>

pub fn parse(bytes: &'a [u8], n: usize, dim: usize) -> Result<Self>

pub fn group_scale(&self, i: usize, g: usize) -> f32

Read the f16 group scale g of row i as f32.

pub fn row_codes(&self, i: usize) -> &'a [u8]

Borrow row i’s packed nibble bytes (dim/2 of them).

pub fn row_scales_f32(&self, i: usize) -> Vec<f32>

Decode row i’s f16 group scales into a fresh Vec<f32> (one per 64-dim block). One-shot use only (the packed ANN store materializes every row once); the per-candidate rerank uses row_scales_into.

pub fn row_scales_into(&self, i: usize, out: &mut [f32])

Decode row i’s f16 group scales into a caller-owned buffer of exactly blocks f32s, so a rerank over thousands of candidates reuses one allocation instead of one Vec per row.

Panics

When out.len() != self.blocks.