Struct QuantConfig 

pub struct QuantConfig {
    pub bits: u8,
    pub per_channel: bool,
    pub symmetric: bool,
    pub calibration_method: Option<CalibrationMethod>,
    pub excluded_layers: Vec<String>,
    pub layer_bits: HashMap<String, u8>,
    pub min_elements: usize,
}

Configuration for a quantization pass.

Fields are public for ergonomic construction in callers; new fields are added only on a minor version bump and downstream ..Default::default() callers continue to compile. Always construct via QuantConfig { /* fields */, ..Default::default() }.

Fields

bits: u8

Bit width: 4 for INT4 or 8 for INT8.

per_channel: bool

When true, compute separate scale/zero-point per output channel (axis 0).

symmetric: bool

When true, force zero_point == 0 (symmetric quantization) — required by most ONNX Runtime / TensorRT INT8 matmul kernels for per-channel weight quantization. Defaults to false (asymmetric).

calibration_method: Option<CalibrationMethod>

Optional calibration method used for range optimization.

excluded_layers: Vec<String>

Layer names to skip entirely (exact match against the initializer name).

layer_bits: HashMap<String, u8>

Per-layer bit-width overrides. Key = initializer name, value = 4 or 8.

min_elements: usize

Minimum number of elements a tensor must have to be quantized. Tensors with fewer elements are left in FP32. Defaults to 0 (no minimum).

Implementations

impl QuantConfig

pub fn int8() -> Self

Create a default INT8 per-tensor configuration.

pub fn with_per_channel(self, enabled: bool) -> Self

Enable or disable per-channel quantization.

pub fn with_symmetric(self, enabled: bool) -> Self

Enable or disable symmetric quantization (zero_point == 0).

pub fn with_calibration(self, method: CalibrationMethod) -> Self

Set the calibration method for range optimization.

pub fn should_quantize(&self, name: &str, num_elements: usize) -> bool

Return true if the layer should be quantized.

A layer is skipped when:

• its name appears in Self::excluded_layers, or
• num_elements is below Self::min_elements (and min_elements > 0).

pub fn bits_for_layer(&self, name: &str) -> u8

Return the effective bit width for a layer.

If the layer name has an entry in Self::layer_bits, that value is used; otherwise the global Self::bits is returned.

Trait Implementations

impl Clone for QuantConfig

fn clone(&self) -> QuantConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantConfig

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

Formats the value using the given formatter.

impl Default for QuantConfig

fn default() -> Self

Returns the “default value” for a type.