Class MxFp8

java.lang.Object

smile.llm.quant.MxFp8

public final class MxFp8
extends Object

MXFP8 (OCP microscaling) weights: FP8 E4M3 elements with one shared E8M0 scale per 32 consecutive elements along the reduction dimension.

An E8M0 byte s encodes the power of two 2^(s - 127); 255 is NaN. A weight W[n][k] is fp8[n][k] * 2^(scale[n][k / 32] - 127). Checkpoints store the scales as <name>.weight_scale_inv (U8, [N, K / 32]).

This is the portable reference: it expands a weight to a dense tensor. Production kernels keep the FP8 bytes and apply the scales in the GEMM.

Field Summary

Fields

Modifier and Type	Field	Description
static final int	BLOCK	Elements sharing one scale.

Method Summary

Modifier and Type	Method	Description
static Tensor	dequantize(Tensor weight, Tensor scale, ScalarType dtype)	Expands an MXFP8 weight to a dense tensor.
static float	e8m0ToFloat(int code)	Returns the value of an E8M0 scale byte.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

BLOCK

public static final int BLOCK

Elements sharing one scale.

See Also:

• Constant Field Values

Method Details

e8m0ToFloat

public static float e8m0ToFloat(int code)

Returns the value of an E8M0 scale byte.

Parameters:

code - the stored byte, 0..255.

Returns:

2^(code - 127), or NaN for 255.

dequantize

public static Tensor dequantize(Tensor weight,
 Tensor scale,
 ScalarType dtype)

Expands an MXFP8 weight to a dense tensor.

Parameters:

weight - FP8 E4M3 weight [N, K] with K a multiple of BLOCK.

scale - E8M0 scales as UInt8 [N, K / 32], on the same device as weight.

dtype - result dtype.

Returns:

the dense weight [N, K]; the caller owns it.

Throws:

IllegalArgumentException - if the shapes are not an MXFP8 layout.