Struct CompressedGradient 

pub struct CompressedGradient {
    pub indices: Vec<usize>,
    pub values: Vec<f32>,
    pub original_size: usize,
    pub compression_ratio: f32,
    pub value_bits: u8,
}

One gradient after compression, in the form it would be put on the wire.

indices is empty for dense methods (quantization, SignSGD): those transmit a value for every coordinate, so an explicit 0..n index list would be pure overhead. Sparse methods (top-k, random-k, threshold) list the coordinates they kept.

values always holds the dequantized f32 values so callers can do arithmetic with them directly; value_bits records how many bits each value actually needs on the wire, which is what CompressedGradient::payload_bytes uses.

Fields

indices: Vec<usize>

Kept coordinates, or empty for a dense positional payload.

values: Vec<f32>

Dequantized values, in indices order (or coordinate order when dense).

original_size: usize

Number of elements in the uncompressed gradient.

compression_ratio: f32

Fraction of the original payload this representation occupies.

value_bits: u8

Bits each value needs on the wire (32 for an untransformed f32, 8 for 8-bit quantization, 1 for a sign).

Implementations

impl CompressedGradient

pub fn payload_bytes(&self) -> usize

Bytes this representation would put on the wire.

Indices cost size_of::<usize>() each; values cost value_bits bits each, rounded up to whole bytes.

pub fn dense_bytes(&self) -> usize

Bytes the uncompressed f32 gradient would occupy.

Trait Implementations

impl Clone for CompressedGradient

fn clone(&self) -> CompressedGradient

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CompressedGradient

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

Formats the value using the given formatter.