Struct ThreadedGradientSync 

pub struct ThreadedGradientSync { /* private fields */ }

All-reduce gradient synchronizer backed by Arc<Mutex<>> for multi-threaded training simulation within a single process.

All workers that share the same underlying SharedState barrier must call GradientSync::sync_gradients with arrays of the same length, otherwise an error is returned. The synchronization algorithm is:

1. Worker adds its gradients into the shared accumulator.
2. The last arriving worker computes the element-wise mean, stores it as the result, and signals all waiters.
3. All workers copy the averaged result back into their local gradient buffer.
4. The last departing worker resets state for the next round.

Implementations

impl ThreadedGradientSync

pub fn new_workers(num_workers: usize) -> Vec<Self>

Create num_workers sync objects that share the same barrier state.

Panics

Panics if num_workers == 0.

pub fn worker_id(&self) -> usize

Return the worker index (0-based) for this instance.

Trait Implementations

impl Clone for ThreadedGradientSync

fn clone(&self) -> ThreadedGradientSync

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ThreadedGradientSync

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

Formats the value using the given formatter.

impl GradientSync for ThreadedGradientSync

fn sync_gradients(&self, gradients: &mut Array1<f32>) -> ModelResult<()>

Synchronize (aggregate) gradients across all workers.

fn is_distributed(&self) -> bool

Returns true if this sync implementation involves multiple workers.

fn num_workers(&self) -> usize

Number of workers participating in synchronization.