Implementation:Huggingface Peft FourierFTConfig

Knowledge Sources	Huggingface_Peft
Domains	Deep_Learning, Parameter_Efficient_Finetuning
Last Updated	2026-02-07 14:00 GMT

Overview

Concrete configuration dataclass for the FourierFT PEFT adapter method provided by the Huggingface PEFT library.

Description

FourierFTConfig controls the hyperparameters for FourierFT, which applies parameter-efficient fine-tuning using Discrete Fourier Transform learnable frequencies. The n_frequency parameter controls the number of trainable parameters per layer, and FourierFT can achieve comparable results to LoRA with significantly fewer parameters. This class extends PeftConfig and supports per-layer frequency patterns.

Usage

Import and use FourierFTConfig when you want to apply Fourier-based parameter-efficient fine-tuning to linear layers, achieving high parameter efficiency compared to LoRA while maintaining competitive accuracy on NLU, instruction tuning, and image classification tasks.

Code Reference

Source Location

• Repository: Huggingface_Peft
• File: src/peft/tuners/fourierft/config.py
• Lines: 24-207

Signature

@dataclass
class FourierFTConfig(PeftConfig):
    n_frequency: int = 1000
    scaling: float = 150.0
    random_loc_seed: Optional[int] = 777
    fan_in_fan_out: bool = False
    target_modules: Optional[Union[list[str], str]] = None
    exclude_modules: Optional[Union[list[str], str]] = None
    bias: str = "none"
    modules_to_save: Optional[list[str]] = None
    layers_to_transform: Optional[Union[list[int], int]] = None
    layers_pattern: Optional[Union[list[str], str]] = None
    n_frequency_pattern: Optional[dict] = field(default_factory=dict)
    init_weights: bool = False

Import

from peft import FourierFTConfig

I/O Contract

Inputs

Name	Type	Default	Description
n_frequency	int	1000	Number of learnable frequencies for DFT. Must be > 0 and <= d^2 for a d x d weight.
scaling	float	150.0	Scaling value for delta W matrix, analogous to lora_alpha in LoRA.
random_loc_seed	Optional[int]	777	Seed for the random location of frequencies (spectral entry matrix).
fan_in_fan_out	bool	False	Set True if layer stores weight as (fan_in, fan_out).
target_modules	Optional[Union[list[str], str]]	None	Module names or regex to apply FourierFT to. Only linear layers supported.
exclude_modules	Optional[Union[list[str], str]]	None	Module names or regex to exclude from FourierFT.
bias	str	"none"	Bias type. Can be 'none', 'all', or 'fourier_only'.
modules_to_save	Optional[list[str]]	None	Additional modules to set as trainable and save in the final checkpoint.
layers_to_transform	Optional[Union[list[int], int]]	None	Specific layer indexes to transform.
layers_pattern	Optional[Union[list[str], str]]	None	Layer pattern name when layers_to_transform is set.
n_frequency_pattern	Optional[dict]	{}	Mapping from layer names or regex to custom n_frequency values per layer.
init_weights	bool	False	If False (default), spectrum initialized to standard normal. If True, initialized to zeros.

Outputs

Name	Type	Description
instance	FourierFTConfig	Configuration object to pass to get_peft_model

Usage Examples

Basic Configuration

from peft import FourierFTConfig, get_peft_model
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("model-name")
config = FourierFTConfig(
    n_frequency=1000,
    scaling=150.0,
    target_modules=["q_proj", "v_proj"],
)
model = get_peft_model(model, config)

Related Pages

• Environment:Huggingface_Peft_Python_Core_Dependencies