Enum ScaleDtype 

pub enum ScaleDtype {
    F32,
    F16,
    BF16,
    UE8M0,
    UE4M3,
}

The data type a scale level stores its scales in.

Variants

F32

Full precision.

F16

Half precision.

BF16

bfloat16 precision.

UE8M0

unsigned floating point, e8m0 format.

UE4M3

unsigned floating point, e4m3 format.

Implementations

impl ScaleDtype

pub const UE8M0_MIN: f32

The smallest and largest values ScaleDtype::UE8M0 represents: 2^-127 and 2^127.

The minimum is subnormal in f32 and the maximum is the largest power of two it holds, so both are spelled as bit patterns rather than computed.

pub const UE8M0_MAX: f32

See ScaleDtype::UE8M0_MIN.

pub fn max_representable(&self) -> f32

The largest finite value representable by the dtype.

A two-level scheme picks its per-tensor scale so that the largest block scale lands here, which is what keeps the block scales inside the range their type can express. That recipe only holds for a block dtype narrower than the scale it divides: dividing by ScaleDtype::F32’s or ScaleDtype::UE8M0’s maximum drives the per-tensor scale subnormal and the renormalized block scales to infinity. A two-level scheme has nothing to gain from those params anyway, since their block scales already reach the full range.

pub fn round_up(&self, scale: f32) -> f32

The smallest value representable by the dtype that is not below scale.

Storing a quantization scale wants this rather than the nearest value. Rounding down puts the scale below what calibration asked for, so every value at the block maximum clips to the quantization range; rounding up costs one step of coarseness instead. Backends have to agree on this, or a tensor quantized on one reconstructs differently on another.

This is not a cast. Conversion to these types rounds to nearest, which is what a cast should do; this is the storage policy for a scale specifically.

scale must not be negative. Symmetric quantization only produces non-negative scales, and the stepping below walks away from zero for a negative input.

ScaleDtype::UE8M0 takes its own path rather than the shared grid below: its range runs to 2^-127, which is subnormal in f32, so the two ends need clamping before the bit stepping is meaningful. Between them the rule is the same one — a ue8m0 value is a bare exponent, so rounding up to it is rounding up to a power of two.

pub fn f32_grid(&self) -> F32Grid

The dtype’s grid, expressed on the f32 bit pattern. See [F32Grid].

bf16 reports no subnormal range because it does not need the separate treatment: its pattern is f32’s top half all the way down, so the bit step stays right where the others stop. Its own subnormals start at 2^-133, which is subnormal in f32 too and flushed to zero by most backends.

Panics

For ScaleDtype::F32, which is the grid itself.

Trait Implementations

impl Clone for ScaleDtype

fn clone(&self) -> ScaleDtype

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for ScaleDtype

impl Debug for ScaleDtype

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for ScaleDtype

fn deserialize<__D>( __deserializer: __D, ) -> Result<ScaleDtype, <__D as Deserializer<'de>>::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Eq for ScaleDtype

impl Hash for ScaleDtype

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

where __H: Hasher,

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 Ord for ScaleDtype

fn cmp(&self, other: &ScaleDtype) -> Ordering

This method returns an Ordering between self and other.

fn max(self, other: Self) -> Self

where Self: Sized,

Compares and returns the maximum of two values.

fn min(self, other: Self) -> Self

where Self: Sized,

Compares and returns the minimum of two values.

fn clamp(self, min: Self, max: Self) -> Self

where Self: Sized,

Restrict a value to a certain interval.

impl PartialEq for ScaleDtype

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl PartialOrd for ScaleDtype

fn partial_cmp(&self, other: &ScaleDtype) -> Option<Ordering>

This method returns an ordering between self and other values if one exists.

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

Tests less than (for self and other) and is used by the < operator.

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

Tests less than or equal to (for self and other) and is used by the <= operator.

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

Tests greater than (for self and other) and is used by the > operator.

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

Tests greater than or equal to (for self and other) and is used by the >= operator.

impl Serialize for ScaleDtype

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

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for ScaleDtype