Module quantization 

Core quantization logic for INT8 and INT4.

Provides tensor-level quantization (per-tensor and per-channel), INT4 bit-packing, and the high-level Quantizer that combines a QuantConfig with optional calibration statistics.

Structs

Int4Range

Marker for INT4 quantization (-8 … 7).

Int8Range

Marker for INT8 quantization (-128 … 127).

QuantConfig

Configuration for a quantization pass.

QuantParamsGeneric

Affine quantization parameters (scale and zero-point), generic over bit-width.

QuantizedTensorGeneric

Generic quantized tensor, parameterized by bit-width marker.

Quantizer

High-level quantizer that combines configuration with optional calibration.

Enums

QuantizedTensorType

Type-erased wrapper over QuantizedTensor (INT8) and QuantizedTensorInt4 (INT4).

Traits

QuantRange

Marker trait that supplies the clamp constants for a quantization bit-width.

Functions

pack_int4

Pack a slice of INT4 values (two per byte, high nibble first).

unpack_int4

Unpack INT4 values from packed bytes, returning exactly num_values i8s.

Type Aliases

QuantParams

INT8 affine quantization parameters — clamp(-128, 127).

QuantParamsInt4

INT4 affine quantization parameters — clamp(-8, 7).

QuantizedTensor

An INT8 quantized tensor with optional per-channel parameters.

QuantizedTensorInt4

An INT4 quantized tensor with optional per-channel parameters and bit packing.