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§
- Client
Selection - How clients are selected per federated round.
- Communication
Protocol - How model-parallel partitions communicate activations and gradients.
- Exploit
Strategy - PBT exploit strategy: how underperformers learn from top performers.
- Explore
Strategy - PBT explore strategy: how hyperparameters are mutated after exploit.
- Federated
Aggregation - Aggregation method for federated learning rounds.
- Gradient
Aggregation - How gradients are aggregated across workers in data-parallel training.
- Training
Strategy - Training strategy — graph-level attribute, inherited by subgraphs.