Ft Libraries

Framework registry for managing fine-tuning libraries.

This module provides access to the registered fine-tuning frameworks and helper functions to get a library instance by name.

FrameworkRegistry

Registry for available fine-tuning frameworks.

get(name, **kwargs) classmethod

Get an instantiated fine-tuning library by name.

Parameters:

Name	Type	Description	Default
name	str	Name of the framework.	required
**kwargs	Any	Additional arguments to pass to the library's constructor (if any).	{}

Returns:

Type	Description
BaseFineTuningLibrary	An instance of the fine-tuning library.

Raises:

Type	Description
ValueError	If the framework name is not recognized.

list_available_frameworks() classmethod

List all registered framework names.

Returns:

Type	Description
list[str]	A list of framework names.

register(name, library_class) classmethod

Register a fine-tuning library class.

Libraries are instantiated on demand.

Parameters:

Name	Type	Description	Default
name	str	Name of the framework (e.g., "unsloth").	required
library_class	Type[BaseFineTuningLibrary]	The class of the fine-tuning library.	required

UnslothFineTuningLibrary()

Bases: UnslothBaseLibrary, FineTuningLibraryProtocol[Union[UnslothModel, UnslothMultimodalModel], SFTTrainer, UnslothTokenizer]

FineTuningLibrary implementation for the Unsloth framework with optimized model preparation.

Initialize UnslothFineTuningLibrary.

Raises:

Type	Description
ImportError	If unsloth is not available (GPU not available or not installed).

framework_name: str property

Get the name of the framework.

Returns:

Name	Type	Description
str	str	The name of the framework.

load_model(experiment_args, hyperparameters, **kwargs)

Load an Unsloth model and initialize its tokenizer.

The tokenizer is stored internally and can be retrieved with get_tokenizer.

Parameters:

Name	Type	Description	Default
experiment_args	ExperimentConfig	Configuration for the experiment.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
**kwargs	Any	Additional arguments (not used by Unsloth library currently).	{}

Returns:

Type	Description
Tuple[Union[UnslothModel, UnslothMultimodalModel], UnslothTokenizer]	The loaded model instance.

Raises:

Type	Description
ImportError	If Unsloth is not installed and is required.
ValueError	If required model arguments are missing.

save_model(trainer, path=None, experiment_args=None, hyperparameters=None, storage_args=None, validator=None, results=None, model_path=None)

Save the trained model (and its tokenizer) to the specified path.

This method supports two calling conventions: 1. Simple: save_model(trainer, path) - for use by utility functions 2. Extended: save_model(trainer, experiment_args=..., hyperparameters=..., ...) - for use by trainers

Parameters:

Name	Type	Description	Default
trainer	Any	The trainer instance (TrlSFTTrainer, TrlDPOTrainer, or TrlGRPOTrainer).	required
path	str | None	The directory path to save the model (simple mode). Defaults to None.	None
experiment_args	ExperimentConfig | None	Experiment configuration (extended mode). Defaults to None.	None
hyperparameters	FinetunedHyperparameters | None	Hyperparameters configuration (extended mode). Defaults to None.	None
storage_args	StorageConfig | None	Storage configuration (extended mode). Defaults to None.	None
validator	Any	Validator instance for model validation (extended mode). Defaults to None.	None
results	dict[str, Any] | None	Training results for newly trained models (extended mode). Defaults to None.	None
model_path	str | None	Path to the model directory for existing models (extended mode). Defaults to None.	None

Returns:

Name	Type	Description
str	str | None	Path where the model was saved or uploaded (extended mode), or None (simple mode).

Raises:

Type	Description
ValueError	If neither model_path nor results is provided in extended mode, or if configuration is invalid.
RuntimeError	If the save operation fails.

save_tokenizer(tokenizer, path, **kwargs)

Saves the tokenizer to the specified path.

Parameters:

Name	Type	Description	Default
tokenizer	UnslothTokenizer	The tokenizer instance to save.	required
path	str	The directory path to save the tokenizer to.	required
**kwargs	Any	Additional framework-specific arguments.	{}

setup_training(model, tokenizer, hyperparameters, train_dataset, eval_dataset=None, is_bfloat16_supported=False, dataset_text_field='prompt', formatting_func=None)

Set up a TRL SFTTrainer for the model.

Parameters:

Name	Type	Description	Default
model	Union[UnslothModel, UnslothMultimodalModel]	The model instance to train (should be from load_model).	required
tokenizer	UnslothTokenizer	The tokenizer instance to use.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
is_bfloat16_supported	bool	Whether bfloat16 is supported by hardware.	False
train_dataset	Any	The training dataset.	required
eval_dataset	Optional[Any]	The evaluation dataset (optional).	None
dataset_text_field	str	Field in dataset containing the text data. Defaults to "prompt".	'prompt'
formatting_func	Callable[[dict], list[str]]	Function to format the dataset.	None

Returns:

Type	Description
SFTTrainer	The initialized TRL SFTTrainer instance.

Raises:

Type	Description
RuntimeError	If the model or tokenizer (passed as arg) is invalid.
ImportError	If Unsloth/TRL is not available.
ValueError	If evaluation strategy is set but no validation dataset is provided.

setup_training_dpo(model, tokenizer, hyperparameters, train_dataset, eval_dataset=None, **kwargs)

Set up a TRL DPOTrainer for the model.

Parameters:

Name	Type	Description	Default
model	Union[UnslothModel, UnslothMultimodalModel]	The model instance to train (should be from load_model).	required
tokenizer	UnslothTokenizer	The tokenizer instance to use.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
train_dataset	Any	The training dataset.	required
eval_dataset	Optional[Any]	The evaluation dataset (optional).	None
**kwargs	Any	Additional arguments.	{}

Returns:

Type	Description
DPOTrainer	The initialized TRL DPOTrainer instance.

Raises:

Type	Description
RuntimeError	If the model or tokenizer (passed as arg) is invalid.
ImportError	If Unsloth/TRL is not available.

setup_training_grpo(model, tokenizer, hyperparameters, vllm_sampling_params_config, train_dataset, eval_dataset=None, reward_funcs=None, **kwargs)

Set up a TRL GRPOTrainer for the model.

Parameters:

Name	Type	Description	Default
model	Union[UnslothModel, UnslothMultimodalModel]	The model instance to train (should be from load_model).	required
tokenizer	UnslothTokenizer	The tokenizer instance to use.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required
vllm_sampling_params_config	VllmSamplingParamsConfig	Configuration for vLLM sampling parameters.	required
train_dataset	Any	The training dataset.	required
eval_dataset	Optional[Any]	The evaluation dataset (optional).	None
reward_funcs	list[Callable[[str], float]]	List of reward functions for GRPO training.	None
**kwargs	Any	Additional arguments.	{}

Returns:

Type	Description
GRPOTrainer	The initialized TRL GRPOTrainer instance.

Raises:

Type	Description
RuntimeError	If the model or tokenizer (passed as arg) is invalid.
ImportError	If Unsloth/TRL is not available.

get_framework_library(name, **kwargs)

High-level function to get a framework library instance.

Parameters:

Name	Type	Description	Default
name	str	The name of the framework.	required
**kwargs	Any	Additional arguments for the library's constructor.	{}

Returns:

Name	Type	Description
BaseFineTuningLibrary	BaseFineTuningLibrary	An instance of the requested library.

list_available_frameworks()

List all available frameworks.

Returns:

Type	Description
list[str]	A list of framework names.

trainer_preparation(experiment_args, hyperparameters)

Create a trainer using the specified framework.

Parameters:

Name	Type	Description	Default
experiment_args	ExperimentConfig	Configuration for the experiment.	required
hyperparameters	FinetunedHyperparameters	Configuration for the hyperparameters.	required

Returns:

Type	Description
Tuple[Any, Any, Any]	Tuple[Any, Any, Any]: A tuple containing: - library: The framework library instance. - model: The loaded model instance. - tokenizer: The tokenizer instance.

Raises:

Type	Description
ValueError	If the framework is not registered, initialization fails, or any other error occurs.