Struct Fp8Codec 

pub struct Fp8Codec {
    pub format: Fp8Format,
    pub saturate: bool,
}

FP8 encoder/decoder.

Saturating conversion: values exceeding max_val are clamped; NaN/Inf inputs return an error.

Fields

format: Fp8Format

Target FP8 format.

saturate: bool

Whether to saturate (clamp) out-of-range values instead of erroring.

Implementations

impl Fp8Codec

pub fn e4m3() -> Self

E4M3 codec with saturation.

pub fn e5m2() -> Self

E5M2 codec with saturation.

pub fn encode_f32(&self, v: f32) -> QuantResult<u8>

Encode a single f32 to FP8 u8.

Errors

• QuantError::NonFiniteFp8 if v is NaN or Inf and saturate = false.

pub fn decode_f32(&self, b: u8) -> f32

Decode a FP8 u8 to f32.

pub fn encode(&self, data: &[f32]) -> QuantResult<Vec<u8>>

Encode a slice of f32 to FP8.

Errors

Propagates QuantError::NonFiniteFp8 on first NaN/Inf when not saturating.

pub fn decode(&self, data: &[u8]) -> Vec<f32>

Decode a slice of FP8 to f32.

pub fn quantization_mse(&self, data: &[f32]) -> QuantResult<f32>

Mean squared error of the FP8 round-trip.

Errors

Propagates encode errors.

Trait Implementations

impl Clone for Fp8Codec

fn clone(&self) -> Fp8Codec

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Fp8Codec

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

Formats the value using the given formatter.

impl Copy for Fp8Codec