Crate peft_rs 

peft-rs

Comprehensive PEFT (Parameter-Efficient Fine-Tuning) adapter library for Rust.

This crate provides modular implementations of various PEFT methods:

• LoRA (Low-Rank Adaptation)
• DoRA (Weight-Decomposed Low-Rank Adaptation)
• AdaLoRA (Adaptive Low-Rank Adaptation)
• IA³ (Infused Adapter by Inhibiting and Amplifying Inner Activations)
• LoHa (Low-Rank Hadamard Product)
• LoKr (Low-Rank Kronecker Product)
• OFT (Orthogonal Fine-Tuning)
• BOFT (Butterfly Orthogonal Fine-Tuning)
• VeRA (Vector-based Random Matrix Adaptation)
• Prefix Tuning
• Prompt Tuning

Quick Start

use peft_rs::{LoraConfig, LoraLayer};
use candle_core::{Device, Tensor};

// Create a LoRA layer
let config = LoraConfig {
    r: 8,
    alpha: 16,
    dropout: 0.0,
};
let layer = LoraLayer::new(768, 768, config, &Device::Cpu)?;

// Apply to input
let input = Tensor::zeros(&[1, 10, 768], candle_core::DType::F32, &Device::Cpu)?;
let output = layer.forward(&input)?;

Architecture

All adapters implement the Adapter trait, which provides a common interface for forward passes and weight merging.

Re-exports

pub use adapters::adalora::AdaLoraConfig;

pub use adapters::adalora::AdaLoraLayer;

pub use adapters::boft::BoftConfig;

pub use adapters::boft::BoftLayer;

pub use adapters::ia3::Ia3Config;

pub use adapters::ia3::Ia3Layer;

pub use adapters::loha::LoHaConfig;

pub use adapters::loha::LoHaLayer;

pub use adapters::lokr::LoKrConfig;

pub use adapters::lokr::LoKrLayer;

pub use adapters::lora::DoraLayer;

pub use adapters::lora::LoraConfig;

pub use adapters::lora::LoraLayer;

pub use adapters::oft::OftConfig;

pub use adapters::oft::OftLayer;

pub use adapters::prefix_tuning::PrefixTuningConfig;

pub use adapters::prefix_tuning::PrefixTuningLayer;

pub use adapters::prompt_tuning::PromptTuningConfig;

pub use adapters::prompt_tuning::PromptTuningLayer;

pub use adapters::vera::VeraConfig;

pub use adapters::vera::VeraLayer;

pub use error::PeftError;

pub use error::Result;

pub use io::load_adapter_config;

pub use io::load_adapter_weights;

pub use io::load_pretrained;

pub use io::save_adapter_config;

pub use io::save_adapter_weights;

pub use io::save_pretrained;

pub use io::SaveLoad;

pub use io::ADAPTER_CONFIG_FILENAME;

pub use io::ADAPTER_WEIGHTS_FILENAME;

pub use model::get_peft_model;

pub use model::ModulePattern;

pub use model::PeftModel;

pub use registry::AdapterRegistry;

pub use training::count_trainable_parameters;

pub use training::format_parameter_count;

pub use training::AdapterTrainingConfig;

pub use training::AdapterTrainingState;

pub use training::LrSchedule;

pub use traits::Adapter;

pub use traits::AdapterConfig;

pub use traits::Mergeable;

pub use traits::Trainable;

Modules

adapters

PEFT adapter implementations.

config

Configuration types for PEFT adapters.

error

Error types for peft-rs.

io

I/O utilities for saving and loading adapter weights and configurations.

model

Model integration for PEFT adapters.

registry

Multi-adapter registry for managing multiple PEFT adapters.

training

Training utilities for PEFT adapters.

traits

Core traits for PEFT adapters.