Parameters that fully describe a quantization mapping. More...

#include <quantized_vector.hpp>

Public Member Functions
std::string	serialize () const
	Serialize to a compact key-value string for Parquet metadata.

size_t	bytes_per_vector () const
	Storage size per vector in bytes.

Static Public Member Functions
static QuantizationParams	compute (const float *data, size_t num_vectors, uint32_t dim, QuantizationScheme scheme)
	Compute optimal quantization parameters from a batch of vectors.

static expected< QuantizationParams >	deserialize (const std::string &s)
	Deserialize from the key-value string produced by serialize().

Public Attributes
QuantizationScheme	scheme
	Quantization scheme (symmetric/asymmetric, INT8/INT4).

float	scale = 0.0f
	Scale factor (float units per quantization step).

float	zero_point = 0.0f
	Offset (used by ASYMMETRIC_INT8 only).

uint32_t	dimension = 0
	Vector dimension (number of float elements).

Detailed Description

Parameters that fully describe a quantization mapping.

Holds the scheme, scale factor, zero point, and vector dimension needed to quantize float32 vectors to INT8/INT4 and dequantize them back.

See also: Quantizer, Dequantizer, QuantizationScheme

Definition at line 76 of file quantized_vector.hpp.

Member Function Documentation

◆ bytes_per_vector()

size_t signet::forge::QuantizationParams::bytes_per_vector ( ) const

inline

Storage size per vector in bytes.

SYMMETRIC_INT8 / ASYMMETRIC_INT8: dim bytes.
SYMMETRIC_INT4: ceil(dim / 2) bytes.

Returns: Number of bytes needed to store one quantized vector.

Definition at line 540 of file quantized_vector.hpp.

◆ compute()

QuantizationParams signet::forge::QuantizationParams::compute	(	const float *	data,
		size_t	num_vectors,
		uint32_t	dim,
		QuantizationScheme	scheme
	)

inlinestatic

Compute optimal quantization parameters from a batch of vectors.

Scans all float values to determine min/max, then derives scale (and zero_point for asymmetric) accordingly.

Parameters

data	Pointer to `num_vectors * dim` contiguous floats.
num_vectors	Number of vectors in the batch.
dim	Dimensionality of each vector.
scheme	Desired quantization scheme.

Returns: QuantizationParams ready for use with Quantizer/Dequantizer.

Definition at line 372 of file quantized_vector.hpp.

◆ deserialize()

expected< QuantizationParams > signet::forge::QuantizationParams::deserialize ( const std::string & s )

inlinestatic

Deserialize from the key-value string produced by serialize().

Parameters

s	The serialized parameter string.

Returns: The parsed QuantizationParams, or an error if malformed.

Definition at line 454 of file quantized_vector.hpp.

◆ serialize()

std::string signet::forge::QuantizationParams::serialize ( ) const

inline

Serialize to a compact key-value string for Parquet metadata.

Format: "scheme=0;scale=0.001234;zero_point=0.0;dimension=768".

Returns: The serialized parameter string.

Definition at line 440 of file quantized_vector.hpp.

Member Data Documentation

◆ dimension

uint32_t signet::forge::QuantizationParams::dimension = 0

Vector dimension (number of float elements).

Definition at line 80 of file quantized_vector.hpp.

◆ scale

float signet::forge::QuantizationParams::scale = 0.0f

Scale factor (float units per quantization step).

Definition at line 78 of file quantized_vector.hpp.

◆ scheme

QuantizationScheme signet::forge::QuantizationParams::scheme

Quantization scheme (symmetric/asymmetric, INT8/INT4).

Definition at line 77 of file quantized_vector.hpp.

◆ zero_point

float signet::forge::QuantizationParams::zero_point = 0.0f

Offset (used by ASYMMETRIC_INT8 only).

Definition at line 79 of file quantized_vector.hpp.

The documentation for this struct was generated from the following file:

include/signet/ai/quantized_vector.hpp

Public Member Functions

Static Public Member Functions

Public Attributes