Struct FloatLayout 

pub struct FloatLayout {
    pub exponent_bits: u8,
    pub mantissa_bits: u8,
    pub specials: Specials,
    pub bias_override: Option<i16>,
    pub sign: SignEncoding,
}

A sign encoding, exponent_bits, mantissa_bits; at most 32 bits wide. An integer is the layout with exponent_bits 0 and a bias_override of 1 - mantissa_bits, which makes its scale exactly 1.

Fields

exponent_bits: u8

Width of the exponent field; 3 for E3M2, 8 for binary32.

mantissa_bits: u8

Width of the stored mantissa field, excluding the implicit leading one; 2 for E3M2, 23 for binary32.

specials: Specials

Policy for the all-ones exponent and for overflow; Specials::None for E3M2.

bias_override: Option<i16>

Replaces the IEEE bias 2^(exponent_bits - 1) - 1 when set; Some(4) on E3M2 shifts its max normal from 28.0 to 14.0.

sign: SignEncoding

How the sign is encoded; Magnitude for IEEE floats, TwosComplement for int8.

Implementations

impl FloatLayout

pub const E5M2: Self

pub const E5M2_SATURATING: Self

E5M2 widths with saturation: the GGUF bf8 storage policy, which keeps infinities out of stored weights.

pub const F16: Self

IEEE 754 binary16: bias 15, inf and nan reserved, max normal 65504.0.

pub const E2M1: Self

pub const E2M2: Self

Proxima cold-storage E2M2: bias 1, no inf or nan, max normal 7.0.

pub const E3M2: Self

OCP MX v1.0 FP6 E3M2: bias 3, no inf or nan, max normal 28.0 (page 12, section 5.3.2, Table 4).

pub const E2M3: Self

pub const BF32: Self

pub const BF16: Self

pub const BF8: Self = Self::E5M2

pub const BF4: Self = Self::E2M1

pub const INT8: Self

pub const UINT8: Self

pub const INT16: Self

pub const UINT16: Self

pub const INT32: Self

pub const UINT32: Self

pub const UINT1: Self

pub const UINT2: Self

pub const UINT3: Self

pub const UINT4: Self

pub const UINT5: Self

pub const UINT6: Self

pub const SIGN_MAGNITUDE_INT8: Self

Sign-magnitude eight bits: range -127..=127 with a negative zero.

pub const fn bias(&self) -> i16

pub const fn width(&self) -> u8

pub fn same_format(&self, other: &Self) -> bool

Whether two layouts are the same format: equal exponent and mantissa widths, equal specials, and equal effective bias. == also compares the raw bias_override, so a layout spelled with Some(15) and one with None differ under == and agree here.

pub const fn packed_bytes( &self, elements: u64, scale_block: Option<(FloatLayout, u64)>, ) -> u64

Bytes elements values occupy packed densely at width() bits each, plus one scale code per block elements when scale_block is Some((scale, block)), rounded up to whole bytes: 32 elements of E3M2 under an f16 scale are 32 * 6 + 16 = 208 bits, 26 bytes.

pub const fn max_normal(&self) -> f32

pub const fn min_normal(&self) -> f32

pub const fn min_subnormal(&self) -> f32

Trait Implementations

impl Clone for FloatLayout

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for FloatLayout

impl Debug for FloatLayout

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for FloatLayout

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Eq for FloatLayout

impl Hash for FloatLayout

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for FloatLayout

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl Serialize for FloatLayout

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for FloatLayout