signet::forge::Quantizer Class Reference

Quantizes float32 vectors to INT8 or INT4 representation. More...

#include <quantized_vector.hpp>

Public Member Functions
	Quantizer (QuantizationParams params)
	Construct a quantizer with the given parameters.
void	quantize (const float *input, uint8_t *output) const
	Quantize a single float32 vector into the output buffer.
std::vector< uint8_t >	quantize_batch (const float *input, size_t num_vectors) const
	Quantize a batch of vectors into a flat buffer of quantized bytes.
const QuantizationParams &	params () const
	Access the quantization parameters.

Detailed Description

Quantizes float32 vectors to INT8 or INT4 representation.

Uses SIMD acceleration (AVX2, SSE2, or NEON) when available, with a scalar fallback. INT4 quantization always uses the scalar path because nibble packing makes SIMD less beneficial.

See also

QuantizationParams, Dequantizer

Definition at line 131 of file quantized_vector.hpp.

Constructor & Destructor Documentation

Quantizer()

signet::forge::Quantizer::Quantizer ( QuantizationParams  params )

inlineexplicit

Construct a quantizer with the given parameters.

The params must have a valid scale (> 0) and dimension (> 0).

Definition at line 135 of file quantized_vector.hpp.

Member Function Documentation

params()

const QuantizationParams & signet::forge::Quantizer::params ( ) const

inline

Access the quantization parameters.

Definition at line 164 of file quantized_vector.hpp.

quantize()

void signet::forge::Quantizer::quantize ( const float *  input, uint8_t *  output  ) const

inline

Quantize a single float32 vector into the output buffer.

• INT8 schemes: output must hold at least dim bytes.
• INT4 scheme: output must hold at least ceil(dim/2) bytes.

Parameters

input	Pointer to dim float32 values.
output	Destination buffer for quantized bytes.

Definition at line 844 of file quantized_vector.hpp.

quantize_batch()

std::vector< uint8_t > signet::forge::Quantizer::quantize_batch ( const float *  input, size_t  num_vectors  ) const

inline

Quantize a batch of vectors into a flat buffer of quantized bytes.

Parameters

input	Pointer to num_vectors * dim contiguous floats.
num_vectors	Number of vectors to quantize.

Returns

A flat byte buffer of size num_vectors * bytes_per_vector().

Definition at line 879 of file quantized_vector.hpp.

---

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

• include/signet/ai/quantized_vector.hpp