Utils

Utility functions and helpers for GLLM training.

This module contains general-purpose utilities used across the GLLM training framework.

Authors

• Alfan Dinda Rahmawan (alfan.d.rahmawan@gdplabs.id)

Reviewer

• Muhammad Afif Al Hawari (muhammad.a.a.hawari@gdplabs.id)

References

NONE

ComponentsValidator(components=None)

Validator for training components and paths.

This class provides validation methods for models, paths, and other components required for training.

Initialize the validator with training components.

Parameters:

Name	Type	Description	Default
components	Optional[Any]	Training components to validate.	None

update_components(components)

Update the components reference.

Parameters:

Name	Type	Description	Default
components	Any	New training components to validate.	required

validate_components_initialized()

Ensure required components are initialized.

Raises:

Type	Description
RuntimeError	If components are not properly initialized.

validate_model_for_saving()

Validate that model is ready for saving.

Raises:

Type	Description
RuntimeError	If model is not ready for saving.

validate_model_path(model_path) staticmethod

Validate that the model path exists and is accessible.

Parameters:

Name	Type	Description	Default
model_path	str	Path to the model directory.	required

Raises:

Type	Description
FileNotFoundError	If the model path doesn't exist.

FileLoggerCallback(log_path='logs/train_steps.jsonl')

Bases: TrainerCallback

Callback for logging training metrics to a JSONL file.

This callback writes training logs to a JSON Lines (JSONL) file format, capturing step-level metrics, epoch information, and other training statistics during the training process. The logs are written incrementally as training progresses.

Attributes:

Name	Type	Description
log_path		Path to the JSONL log file where training metrics will be written.
f		File handle for the log file (opened during training, closed at end).

Initialize FileLoggerCallback.

Parameters:

Name	Type	Description	Default
log_path		Path to the JSONL file where training logs will be written. Defaults to "logs/train_steps.jsonl". The directory will be created if it does not exist.	'logs/train_steps.jsonl'

on_log(args, state, control, logs=None, **kwargs)

Called every logging_steps during training.

Writes the current training metrics to the log file in JSONL format. Each log entry includes the global step, epoch, and all metrics from the logs dictionary.

Parameters:

Name	Type	Description	Default
args		Training arguments.	required
state		Training state containing global_step and epoch information.	required
control		Training control object.	required
logs		Dictionary containing training metrics to log. Defaults to None.	None
**kwargs		Additional keyword arguments.	{}

on_train_begin(args, state, control, **kwargs)

Called when training begins.

Opens the log file in append mode for writing training metrics.

Parameters:

Name	Type	Description	Default
args		Training arguments.	required
state		Training state.	required
control		Training control object.	required
**kwargs		Additional keyword arguments.	{}

on_train_end(args, state, control, **kwargs)

Called when training ends.

Closes the log file handle if it was opened.

Parameters:

Name	Type	Description	Default
args		Training arguments.	required
state		Training state.	required
control		Training control object.	required
**kwargs		Additional keyword arguments.	{}