Implementation:Huggingface Trl Get Peft Config SFT

Knowledge Sources	LoRA TRL TRL Docs
Domains	NLP, Training
Last Updated	2026-02-06 17:00 GMT

Overview

Concrete wrapper for translating TRL's ModelConfig LoRA settings into a PEFT LoraConfig object, provided by the TRL library.

Description

The get_peft_config() function is a thin adapter between TRL's configuration system and the PEFT library. It reads LoRA-related fields from a ModelConfig dataclass and constructs a LoraConfig object that the SFTTrainer uses to wrap the base model with trainable LoRA adapters via peft.get_peft_model().

If model_args.use_peft is False, the function returns None, indicating that full fine-tuning should be used instead. If use_peft is True but the PEFT library is not installed, it raises a ValueError with installation instructions.

Usage

Use this function in any TRL training script where you want to optionally enable LoRA-based parameter-efficient fine-tuning based on command-line or YAML configuration.

Code Reference

Source Location

Repository: TRL
File: trl/trainer/utils.py (lines 309-332)
File: trl/trainer/model_config.py (lines 18-189, ModelConfig dataclass)

Signature

def get_peft_config(model_args: ModelConfig) -> "PeftConfig | None":
    if model_args.use_peft is False:
        return None

    if not is_peft_available():
        raise ValueError(
            "You need to have PEFT library installed in your environment, "
            "make sure to install `peft`. Make sure to run `pip install -U peft`."
        )

    peft_config = LoraConfig(
        task_type=model_args.lora_task_type,
        r=model_args.lora_r,
        target_modules=model_args.lora_target_modules,
        target_parameters=model_args.lora_target_parameters,
        lora_alpha=model_args.lora_alpha,
        lora_dropout=model_args.lora_dropout,
        bias="none",
        use_rslora=model_args.use_rslora,
        use_dora=model_args.use_dora,
        modules_to_save=model_args.lora_modules_to_save,
    )

    return peft_config

Import

from trl.trainer.utils import get_peft_config
from trl import ModelConfig

I/O Contract

Inputs

Name	Type	Required	Description
model_args	`ModelConfig`	Yes	Configuration dataclass containing all model and LoRA settings
model_args.use_peft	`bool`	Yes	Master switch: if False, returns None (no PEFT)
model_args.lora_r	`int`	No	Rank of LoRA decomposition; default: 16
model_args.lora_alpha	`int`	No	LoRA scaling factor; default: 32
model_args.lora_dropout	`float`	No	Dropout on adapter inputs; default: 0.05
model_args.lora_target_modules	None	No	Layers to inject adapters; None uses PEFT defaults
model_args.lora_target_parameters	None	No	Specific parameter names to target for LoRA
model_args.lora_task_type	`str`	No	PEFT task type; default: `"CAUSAL_LM"`
model_args.use_rslora	`bool`	No	Use rank-stabilized scaling (`alpha/sqrt(r)`); default: False
model_args.use_dora	`bool`	No	Use Weight-Decomposed LoRA; default: False
model_args.lora_modules_to_save	None	No	Additional modules to fully unfreeze and train

Outputs

Name	Type	Description
peft_config	None	A PEFT LoraConfig ready to pass to `SFTTrainer(peft_config=...)`, or None if PEFT is disabled

Usage Examples

Basic Usage

from trl import ModelConfig
from trl.trainer.utils import get_peft_config

model_args = ModelConfig(
    model_name_or_path="Qwen/Qwen2-0.5B",
    use_peft=True,
    lora_r=32,
    lora_alpha=16,
    lora_dropout=0.1,
)

peft_config = get_peft_config(model_args)
print(peft_config)
# LoraConfig(task_type='CAUSAL_LM', r=32, lora_alpha=16, lora_dropout=0.1, bias='none')

Passing to SFTTrainer

from trl import SFTTrainer, SFTConfig, ModelConfig
from trl.trainer.utils import get_peft_config
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-0.5B")
model_args = ModelConfig(use_peft=True, lora_r=16, lora_alpha=32)

trainer = SFTTrainer(
    model=model,
    args=SFTConfig(output_dir="./output"),
    train_dataset=train_dataset,
    peft_config=get_peft_config(model_args),
)

Full Fine-Tuning (No PEFT)

from trl import ModelConfig
from trl.trainer.utils import get_peft_config

model_args = ModelConfig(
    model_name_or_path="Qwen/Qwen2-0.5B",
    use_peft=False,
)

peft_config = get_peft_config(model_args)
assert peft_config is None  # No adapter will be applied

Related Pages

Implements Principle

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment