Struct ParameterServer 

pub struct ParameterServer { /* private fields */ }

Centralized parameter server: rank 0 collects, averages, and broadcasts gradients (or parameters).

Use this instead of AllReduceAggregator when you want a star topology (all workers communicate only with rank 0). This is simpler to reason about and works well when the coordinator has high bandwidth, but can become a bottleneck at large scale compared to ring all-reduce.

• broadcast_params – rank 0 sends the current parameters to all workers (useful at initialisation or after a checkpoint restore).
• reduce_gradients – workers send local gradients to rank 0; rank 0 averages them and broadcasts the result back.

Implementations

impl ParameterServer

pub fn new(config: DistributedConfig, transport: Box<dyn Transport>) -> Self

pub fn broadcast_params( &self, params: &[Tensor], ) -> Result<Vec<Tensor>, ModelError>

Rank 0 broadcasts params to all workers; workers receive and return them.

pub fn reduce_gradients( &self, grads: &[Tensor], ) -> Result<Vec<Tensor>, ModelError>

Workers send gradients to rank 0; rank 0 averages and returns result.