Struct RingAllReduce 

pub struct RingAllReduce { /* private fields */ }

Bandwidth-optimal ring all-reduce / all-gather driver.

A single RingAllReduce simulates the whole ring in-process: each public method takes the per-rank inputs, drives all world_size ranks through the collective using a LocalTransport, and returns each rank’s result (all ranks end up holding identical data).

Implementations

impl RingAllReduce

pub fn new(world_size: usize) -> Result<Self>

Create a driver for a ring of world_size workers.

Returns an error when world_size is zero.

pub fn world_size(&self) -> usize

Number of workers in the ring.

pub fn all_reduce_all<A>( &self, inputs: &[Array1<A>], op: ReduceOp, ) -> Result<Vec<Array1<A>>>

where A: Float + ScalarOperand + Debug,

Run the full simulated ring all-reduce across every rank at once.

inputs[r] is rank r’s vector; all vectors must be non-empty and share the same length. The returned vector holds each rank’s result, which are all identical and equal to the elementwise reduction under op.

pub fn all_gather<A>(&self, inputs: &[Array1<A>]) -> Result<Vec<Array1<A>>>

where A: Float + ScalarOperand + Debug,

Run the simulated ring all-gather across every rank at once.

inputs[r] is rank r’s contribution; contributions may differ in length (this is a gather, not a reduction) but each must be non-empty. The result holds, for every rank, the concatenation inputs[0] || inputs[1] || ... || inputs[N - 1] in rank order.

Trait Implementations

impl Clone for RingAllReduce

fn clone(&self) -> RingAllReduce

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for RingAllReduce

impl Debug for RingAllReduce

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

Formats the value using the given formatter.