For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

FP6Format

struct FP6Format

Selects between the two OCP MX FP6 element encodings.

The numeric parameters returned by the accessors below are the entire difference between the two formats; every routine in this module is parameterized on this type rather than duplicated per encoding.

Implemented traits

AnyType, Copyable, Deinitable, Equatable, ImplicitlyCopyable, Movable, RegisterPassable, TrivialRegisterPassable

comptime members

E2M3

comptime E2M3 = FP6Format(Int(0))

E3M2

comptime E3M2 = FP6Format(Int(1))

Methods

__init__

def __init__(value: Int) -> Self

__eq__

def __eq__(self, other: Self) -> Bool

Returns:

Bool

mantissa_width

def mantissa_width(self) -> Int

Returns the mantissa field width in bits.

Returns:

Int

exponent_width

def exponent_width(self) -> Int

Returns the exponent field width in bits.

Returns:

Int

exponent_bias

def exponent_bias(self) -> Int

Returns the exponent bias.

Returns:

Int

max_exponent

def max_exponent(self) -> Int

Returns the unbiased exponent of the largest finite value.

E2M3 tops out at 7.5 == 1.875 * 2^2 and E3M2 at 28 == 1.75 * 2^4, so this is 2 and 4 respectively. compute_mxfp6_even_scale subtracts it from the block maximum's exponent to derive the E8M0 scale.

Returns:

Int

max_value

def max_value(self) -> Float32

Returns the largest finite magnitude the encoding represents.

Returns:

Float32