MXFP4Quantizer<T, TInput, TOutput>

Models & Types · AiDotNet.Deployment.Optimization.Quantization.Formats

MXFP4 (Microscaling FP4) quantizer - uses shared exponents for efficient 4-bit floating point. Part of the OCP (Open Compute Project) Microscaling specification.

Implements: IQuantizer<T, TInput, TOutput>

For Beginners

MXFP4 is a 4-bit floating point format where groups of numbers share a common scale (exponent). This allows better representation of values across different magnitudes while staying compact.

How It Works

How It Works:

Format Details:

Key Features:

Reference: OCP Microscaling Formats Specification (2023)

Constructors

Constructor Summary
MXFP4Quantizer(QuantizationConfiguration config, int blockSize) Initializes a new instance of the MXFP4Quantizer.

Properties

Property Summary
QuantizationMode Mode { get; }
bool IsCalibrated { get; set; } Gets whether the quantizer has been calibrated.
int BitWidth { get; }
int BlockSize { get; } Gets the block size used for shared scaling.

Methods

Method Summary
IFullModel<T, TInput, TOutput> Quantize(IFullModel<T, TInput, TOutput> model, QuantizationConfiguration config)
double DecodeFromMXFP4(int encoded) Decodes a 4-bit MXFP4 representation to its value.
double GetScaleFactor(string layerName)
int EncodeToMXFP4(double value) Encodes a value to its 4-bit MXFP4 representation.
int GetZeroPoint(string layerName)
void Calibrate(IFullModel<T, TInput, TOutput> model, IEnumerable<TInput> calibrationData)