Struct GradientAggregator 

pub struct GradientAggregator;

Aggregates gradients from distributed workers.

Implementations

impl GradientAggregator

pub fn new() -> Self

Create a new GradientAggregator.

pub fn aggregate_gradients( &self, local_grad: &[f64], strategy: &AllReduceStrategy, ) -> Vec<f64>

Aggregate local_grad from this worker together with gradients that have already been reduced on other workers, using the given strategy.

For the single-worker case the function simply returns a normalised copy of local_grad. In a real multi-node scenario the caller would pass the collected per-worker slices through ring_all_reduce or the tree variant.

pub fn ring_all_reduce(&self, gradients: Vec<Vec<f64>>) -> Vec<f64>

Simulate ring all-reduce over a set of per-worker gradient vectors.

Ring all-reduce arranges n workers in a ring. It runs in two phases:

1. Scatter-reduce (n−1 steps): at step s, each worker w passes the accumulated data for chunk (w − s) mod n to its right neighbour (w + 1) mod n, which adds it to its own copy. After n−1 steps, worker w holds the fully-reduced sum for chunk (w + 1) mod n.

2. All-gather: collect the fully-reduced chunk from each owning worker and divide by n to obtain the mean.

The mathematical result equals the element-wise mean of all input vectors. This simulation runs synchronously on the calling thread with no I/O.

Trait Implementations

impl Clone for GradientAggregator

fn clone(&self) -> GradientAggregator

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GradientAggregator

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

Formats the value using the given formatter.

impl Default for GradientAggregator

fn default() -> GradientAggregator

Returns the “default value” for a type.