Struct Quantizer 

pub struct Quantizer { /* private fields */ }

High-level quantizer that combines configuration with optional calibration.

Implementations

impl Quantizer

pub fn new(config: QuantConfig) -> Self

Create a quantizer with the given configuration (no calibration).

pub fn with_calibration( config: QuantConfig, stats: HashMap<String, ActivationStats>, ) -> Self

Create a quantizer with configuration and pre-collected activation statistics.

pub fn quantize_tensor_with_name( &self, name: &str, data: &[f32], shape: Vec<usize>, ) -> Result<QuantizedTensorType>

Quantize a tensor with optional calibration.

pub fn quantize_tensor( &self, data: &[f32], shape: Vec<usize>, ) -> Result<QuantizedTensorType>

Quantize a tensor using the configured bit width and per-channel setting.

Errors

Returns QuantizeError::InvalidTensor or QuantizeError::UnsupportedConfig.

pub fn quantize_model( &self, model: &OnnxModel, ) -> Result<Vec<QuantizedWeightOutput>>

Quantize every weight in model that passes QuantConfig::should_quantize. Honours per-layer bit-width overrides.

When this quantizer was built with calibration, activation-based range optimization is used for the default bit-width; layers whose bit-width is overridden fall back to weight-only quantization (the calibration stats are keyed by the default configuration).

Skipped weights do not appear in the returned vector.

Trait Implementations

impl Debug for Quantizer

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

Formats the value using the given formatter.