Federated Learning with LoRA (FedLoRA)

This guide explains how to perform federated learning with LoRA (FedLoRA) fine-tuning using the Flower framework in FinLoRA.

Setup and Installation

FedLoRA is implemented using the Flower framework, which provides a flexible platform for federated learning.

Navigate to FedLoRA Directory

First, navigate to the FedLoRA directory:

cd lora/flowertune-llm

Install Dependencies

Install the required dependencies for FedLoRA:

pip install -e .

This will install the flowertune-llm package and its dependencies, including the Flower framework.

Running FedLoRA

Basic FedLoRA Simulation

Run the FedLoRA simulation with default settings:

flwr run .

This command will start a FedLoRA simulation using the default configuration.

Custom Configuration Options

You can customize the FedLoRA configuration using various run-config parameters:

Using Different Base Models:

# Use OpenLLaMA-7B instead of 3B with 8-bit quantization
flwr run . --run-config "model.name='openlm-research/open_llama_7b_v2' model.quantization=8"

Adjusting Training Parameters:

# Run for 50 rounds with 25% client participation
flwr run . --run-config "num-server-rounds=50 strategy.fraction-fit=0.25"

Combining Multiple Configuration Options:

# Custom model with extended training
flwr run . --run-config "model.name='meta-llama/Llama-3.1-8B-Instruct' model.quantization=8 num-server-rounds=30 strategy.fraction-fit=0.3"

Configuration Parameters

The following table describes the key configuration parameters for FedLoRA:

Table 3 FedLoRA Configuration Parameters
Parameter	Description	Default Value	Example Values
`model.name`	Base model to use for federated training	`openlm-research/open_llama_3b_v2`	`meta-llama/Llama-3.1-8B-Instruct`, `openlm-research/open_llama_7b_v2`
`model.quantization`	Quantization bits for the model	`4`	`4`, `8`, `16`
`num-server-rounds`	Number of FedLoRA rounds	`10`	`10`, `20`, `50`
`strategy.fraction-fit`	Fraction of clients participating in each round	`0.1`	`0.1`, `0.25`, `0.5`
`strategy.fraction-evaluate`	Fraction of clients used for evaluation	`0.1`	`0.1`, `0.2`