TrainingConfig¶

Module: fast_llm.engine.training.config

Fields¶

checkpoint — core

Type: TrainingCheckpointConfig Default: (sub-fields optional)

Configuration for checkpoints.

evaluators — core

Type: dict[str, EvaluatorConfig] Default: dict()

A dictionary of evaluation dataset names and their configurations for the validation phase.

export — core

Type: TrainingExportConfig Default: (sub-fields optional)

Configuration for exports.

logs — core

Type: MetricsLogsConfig Default: (sub-fields optional)

Configuration for metric logging.

shutdown — core

Type: ShutdownConfig Default: (sub-fields optional)

Configuration for automated shutdown.

train_iters — core

Type: int Default: 0

Total number of training iterations.

wandb — core

Type: WandbConfig Default: (sub-fields optional)

Configuration for Wandb.

num_workers — optional

Type: int Default: 2

Number of data loading processes for each data iterator.

timeout — feature

Type: float or None Default: 3600

Timeout for lengthy operations such as checkpoint saving and loading, and dataset preparation and sampling.

prefetch_factor — performance

Type: int or None Default: None

Prefetch factor for the data loaders, i.e., number of micro-batches that each worker may prepare in advance.