Struct CalibrationStats 

pub struct CalibrationStats {
    pub absmax: f32,
    pub mean: Vec<f32>,
    pub n_samples: usize,
    pub dims: usize,
    /* private fields */
}

Calibration statistics for quantization

Following Kaizen: continuously improved from query distribution. Uses Welford’s algorithm for numerical stability (Higham, 2002).

Fields

absmax: f32

Maximum absolute value across calibration set

mean: Vec<f32>

Running mean for each dimension

n_samples: usize

Number of samples seen

dims: usize

Embedding dimensions

Implementations

impl CalibrationStats

pub fn new(dims: usize) -> Self

Create new calibration stats for given dimensions

pub fn update(&mut self, embedding: &[f32]) -> Result<(), QuantizationError>

Update calibration with new embedding (Kaizen loop)

Uses Welford’s online algorithm for numerical stability.

pub fn update_batch( &mut self, embeddings: &[Vec<f32>], ) -> Result<(), QuantizationError>

Update with batch of embeddings (Heijunka batching)

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

Get variance for dimension i

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

Get standard deviation for dimension i

pub fn to_quant_params(&self) -> Result<QuantizationParams, QuantizationError>

Convert calibration to quantization parameters

pub fn is_sufficient(&self, min_samples: usize) -> bool

Check if calibration has sufficient samples

Trait Implementations

impl Clone for CalibrationStats

fn clone(&self) -> CalibrationStats

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CalibrationStats

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

Formats the value using the given formatter.