Module backward_pass 

Backward pass coordination via TensorSwap.

This module provides the pure-data structures that coordinate distributed backward-pass gradient streaming. No async or network I/O lives here; the transport layer calls these types as CIDs arrive.

Structs

BackwardPassConfig

Tuning knobs for a BackwardPassCoordinator

BackwardPassCoordinator

Coordinates backward pass gradient streaming via TensorSwap.

BackwardPassStats

Runtime statistics snapshot for a BackwardPassCoordinator

BackwardPassStep

Tracks one layer’s backward pass across multiple peers

Enums

AggregationMethod

Aggregation strategy for combining gradients from multiple peers

BackwardStepStatus

Coordination status for a distributed backward pass step

Functions

clip_gradient_norm

Clip a gradient vector in-place so that its L2 norm does not exceed max_norm.

federated_average

Compute the unweighted mean of a collection of gradient vectors.