Struct QuantParams 

pub struct QuantParams {
    pub scale: f64,
    pub zero_point: i32,
    pub qmin: i32,
    pub qmax: i32,
}

Affine/symmetric integer quantization parameters.

Holds the grid step (scale), the integer zero_point (always 0 for the symmetric scheme) and the integer clamp range [qmin, qmax].

Fields

scale: f64

Grid step: the real spacing between adjacent integer codes.

zero_point: i32

Integer code onto which the real value 0 maps.

qmin: i32

Lowest representable integer code.

qmax: i32

Highest representable integer code.

Implementations

impl QuantParams

pub fn from_minmax( min_v: f64, max_v: f64, dtype: IntDtype, scheme: QuantScheme, ) -> Result<Self, GpuOptimError>

Build parameters directly from a real [min, max] interval.

For QuantScheme::Symmetric this reduces to Self::from_absmax over max(|min|, |max|). For QuantScheme::Affine the interval is nudged to include the real 0, scale = (max - min) / (qmax - qmin) and the zero_point is chosen so the real 0 maps onto an exact integer code.

Errors

Returns GpuOptimError::InvalidState if the inputs are non-finite, inverted (max < min), or describe a degenerate zero-width interval.

pub fn from_absmax(absmax: f64, dtype: IntDtype) -> Result<Self, GpuOptimError>

Build symmetric parameters from an absolute maximum.

scale = absmax / qmax, zero_point = 0.

Errors

Returns GpuOptimError::InvalidState when absmax is non-finite or not strictly positive (a degenerate, all-zero tensor cannot be calibrated).

pub fn per_tensor_minmax( x: &Array1<f64>, dtype: IntDtype, scheme: QuantScheme, ) -> Result<Self, GpuOptimError>

Calibrate per-tensor parameters from the exact [min, max] of x.

Errors

Returns GpuOptimError::InvalidState for an empty tensor or a degenerate (zero-range) calibration.

pub fn symmetric_from_absmax( x: &Array1<f64>, dtype: IntDtype, ) -> Result<Self, GpuOptimError>

Calibrate symmetric parameters from max(|x|).

Errors

Returns GpuOptimError::InvalidState for an empty or all-zero tensor.

pub fn per_tensor_percentile( x: &Array1<f64>, dtype: IntDtype, scheme: QuantScheme, clip_fraction: f64, ) -> Result<Self, GpuOptimError>

Calibrate per-tensor parameters after clipping the tails.

The interval is taken between the clip_fraction and 1 - clip_fraction quantiles (nearest-rank) of x, which suppresses outliers before the scale is derived.

Errors

Returns GpuOptimError::InvalidState for an empty tensor, a clip_fraction outside [0, 0.5), or a degenerate calibration.

pub fn quantize(&self, x: f64, mode: RoundingMode, rng: &mut impl Rng) -> i32

Quantize a single real value to an integer code (with clamping).

pub fn dequantize(&self, q: i32) -> f64

Dequantize an integer code back to a real value.

pub fn fake_quant_scalar( &self, x: f64, mode: RoundingMode, rng: &mut impl Rng, ) -> f64

Full fake-quant round trip for a single value, dequant(quant(x)).

pub fn real_min(&self) -> f64

Lowest real value representable before the clamp saturates, qmin*scale after removing the zero-point.

pub fn real_max(&self) -> f64

Highest real value representable before the clamp saturates.

Trait Implementations

impl Clone for QuantParams

fn clone(&self) -> QuantParams

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for QuantParams

impl Debug for QuantParams

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

Formats the value using the given formatter.

impl PartialEq for QuantParams

fn eq(&self, other: &QuantParams) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for QuantParams