Skip to main content

Module strategy

Module strategy 

Source
Expand description

Training strategies for distributed execution.

A TrainingStrategy is a graph-level attribute that controls HOW the Scheduler distributes work across workers and HOW workers coordinate during training (gradient aggregation, state sync, communication).

Only the description lives here. Running one — sharding inputs, calling workers in a round loop, aggregating gradients — is execution, and is in somatize_runtime::strategy along with the traits that describe it.

Subgraphs inherit the parent’s strategy unless overridden.

Structs§

Partition
A partition maps a set of node IDs to a worker target.

Enums§

ClientSelection
How clients are selected per federated round.
CommunicationProtocol
How model-parallel partitions communicate activations and gradients.
ExploitStrategy
PBT exploit strategy: how underperformers learn from top performers.
ExploreStrategy
PBT explore strategy: how hyperparameters are mutated after exploit.
FederatedAggregation
Aggregation method for federated learning rounds.
GradientAggregation
How gradients are aggregated across workers in data-parallel training.
TrainingStrategy
Training strategy — graph-level attribute, inherited by subgraphs.