Struct SmoothQuantCalibrator 

pub struct SmoothQuantCalibrator { /* private fields */ }

Online per-channel activation calibrator for SmoothQuant.

Feed batches of activations for each named linear layer via record_activation, then call smooth_factors to obtain the SmoothQuant smoothing vector for that layer’s weight matrix.

Implementations

impl SmoothQuantCalibrator

pub fn new(config: SmoothQuantConfig) -> Self

Create a new calibrator using the given SmoothQuant config.

pub fn record_activation( &mut self, layer_name: &str, activations: &[f32], in_features: usize, )

Record one batch of activations for a named layer.

activations is a [num_tokens × in_features] row-major flat slice. If this is the first call for layer_name, a new per-channel accumulator is created. For subsequent calls the running per-channel max-abs is updated.

Panics

Panics if in_features changes between calls for the same layer_name.

pub fn smooth_factors( &self, layer_name: &str, weights: &[f32], out_features: usize, ) -> Result<Vec<f32>, SmoothQuantError>

Compute SmoothQuant smoothing factors for a named layer.

Uses the running per-channel activation max accumulated via record_activation together with the supplied weight matrix to derive per-input-feature smoothing factors.

weights is [out_features × in_features] row-major.

Returns a Vec<f32> of length in_features.

pub fn layer_count(&self) -> usize

Number of distinct layers recorded by this calibrator.

pub fn has_layer(&self, name: &str) -> bool

Whether the calibrator has any recorded data for the given layer name.