Struct QuantizedTensor 

pub struct QuantizedTensor {
    pub data: Vec<i8>,
    pub scales: Vec<f64>,
    pub zero_points: Vec<f64>,
    pub shape: Vec<usize>,
    pub original_dtype: DType,
    pub config: QuantConfig,
}

A quantized tensor storing integer weights with associated scale/zero_point.

The original float value is recovered by: float = (int - zero_point) * scale

For per-channel quantization, scales and zero_points have one entry per channel (dimension 0 of the original tensor).

Fields

data: Vec<i8>

Quantized integer values (stored as i8 for INT8, packed for INT4).

scales: Vec<f64>

Scale factor(s). Length 1 for per-tensor, N for per-channel.

zero_points: Vec<f64>

Zero point(s). Length 1 for per-tensor, N for per-channel.

shape: Vec<usize>

Original tensor shape.

original_dtype: DType

Original dtype (for dequantization target).

config: QuantConfig

Quantization config used.

Implementations

impl QuantizedTensor

pub fn numel(&self) -> usize

Total number of elements.

pub fn size_bytes(&self) -> usize

Size in bytes of the quantized representation.

pub fn compression_ratio(&self) -> f64

Compression ratio vs FP32.

Trait Implementations

impl Clone for QuantizedTensor

fn clone(&self) -> QuantizedTensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizedTensor

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

Formats the value using the given formatter.