Struct GradientAccumulator 

pub struct GradientAccumulator {
    pub dimension: usize,
    pub clip_strategy: ClipStrategy,
    pub pending: Vec<PeerGradient>,
    pub stats: AccumulatorStats,
}

Accumulates gradients from multiple federated peers and produces a weighted-averaged, optionally-clipped aggregate on each commit.

Fields

dimension: usize

Expected length of every gradient vector.

clip_strategy: ClipStrategy

Clipping strategy applied after averaging.

pending: Vec<PeerGradient>

Gradients waiting to be committed.

stats: AccumulatorStats

Running statistics updated on each successful commit.

Implementations

impl GradientAccumulator

pub fn new(dimension: usize, clip_strategy: ClipStrategy) -> Self

Create a new accumulator expecting gradients of length dimension.

pub fn add(&mut self, pg: PeerGradient) -> Result<(), AccumulationError>

Add a peer gradient to the pending queue.

Returns AccumulationError::DimensionMismatch when pg.gradient.len() != self.dimension.

pub fn commit(&mut self) -> Result<Vec<f32>, AccumulationError>

Commit all pending gradients, producing a weighted average.

After a successful commit the pending queue is cleared and AccumulatorStats is updated.

Errors

• AccumulationError::EmptyAccumulator — no pending gradients.
• AccumulationError::WeightSumZero — all weights are 0.0.

pub fn apply_clip(&self, grad: &mut [f32]) -> bool

Apply the accumulator’s ClipStrategy to grad in-place.

Returns true if any modification was made.

pub fn pending_count(&self) -> usize

Number of gradients currently waiting to be committed.

pub fn stats(&self) -> &AccumulatorStats

Borrow the current statistics.

pub fn reset(&mut self)

Clear all pending gradients and reset statistics to defaults.

Trait Implementations

impl Debug for GradientAccumulator

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

Formats the value using the given formatter.