TrainingExportConfig¶

Module: fast_llm.engine.training.config

Inherits from: TrainingCheckpointBaseConfig, IntervalConfig, CheckpointStateSaveConfigBase

Fields¶

callback — core

Type: CallbackConfig Default: (sub-fields optional)

Callback (shell script) to run after export.

format — core

Type: type Default: FastLLMCheckpointFormat

Format of the checkpoint.

data_type — feature

Type: DataType or None Default: None

Data type to save the checkpoint.

interval — feature

Type: int or None Default: None

The number of training iterations between each export. Setting to None will disable exports.

keep — feature

Type: int or None Default: None

The maximum number of saves to keep. When exceeding this value, checkpoints are deleted starting from the older ones.

keep_every — feature

Type: int or None Default: None

Keep every nth saves, i.e. Exclude it from the checkpoint count and deletion in keep.

model_weights — feature

Type: bool Default: True

Save the model weights.

offset — feature

Type: int Default: 0

Offset for the first export.

optimizer_state — feature

Type: bool Default: None

Save the optimizer state. Default: save if supported by the format.

parameters_per_file — feature

Type: int Default: 4_294_967_296

Limit the number of parameters saved in each file.