Introducing OpenFL 1.6: Federated LLM Fine-Tuning and Evaluation

The OpenFL team is excited to announce the release of OpenFL 1.6, which delivers several new features as well as API and documentation enhancements, enabling the next wave of federated learning experiments. OpenFL 1.6 is a game-changer for developers working with Large Language Models (LLMs), offering a suite of new features specifically designed to help them transition from a centralized to a federated setting.

OpenFL 1.6 focuses on improving the developer experience by streamlining APIs and providing more examples of using OpenFL in emerging scenarios such as federated evaluation and federated fine-tuning of LLMs. This new release further demonstrates OpenFL’s growing support for modern computing frameworks and AI accelerators, ensuring that federated training and inference can be conducted with the highest performance and scalability.

This release also continues the security hardening efforts as a foundation for trusted federated learning, including fixes for static code analysis issues and updates to third-party libraries. To help address more advanced security threats, there are code samples and recipes for protecting federations against model poisoning attacks.

What’s New in OpenFL 1.6?

OpenFL 1.6 is packed with updates that cater to the evolving needs of the federated learning community, with a particular emphasis on LLMs. Here’s what you can look forward to in this release:

• Federated LLM Fine-Tuning
   Learn how to fine-tune neuralchat-7b using the Intel® Extension for Transformers and the OpenFL Workflow API. Or use horovod to efficiently train LLMs across multiple private clusters.
• Workflow API Enhancements
  Introducing an experimental Workspace Export feature that can be used to transform a Workflow API-based FL experiment into the TaskRunner API format for running in a distributed deployment. There is also groundwork laid for a future FederatedRuntime implementation for Workflow API, in addition to the currently supported LocalRuntime.
• Federated Evaluation
  Federated evaluation allows for the assessment of ML models in a federated learning system by validating the model’s performance locally on decentralized collaborator nodes, and then aggregating these metrics to gauge overall effectiveness, without compromising data privacy and security. Federated evaluation is now officially supported by OpenFL, including example tutorials on how to use this new feature, using the TaskRunner API.
• Expanded AI Accelerator Support
  Get started with XPU (Intel® Data Center GPU Max Series) using the Intel® Extension for PyTorch, including examples for training on datasets such as MNIST (using the Workflow API) and Tiny ImageNet (using the Interactive API).  With the addition of a RayExecutor as a backend for the Workflow API’s LocalRuntime, you’ll benefit from Ray’s GPU management capabilities, and ability to scale computation over multiple nodes. This is further illustrated through an example notebook for GPU-accelerated model validation on the aggregator.
• Improved Straggler Collaborator Handling
  Improvements and bug fixes to the aggregator fault-tolerance when collaborators stop responding or drop out of a federation. Introducing a cut-off timer-based policy and enabling other policies to be plugged-in. This capability is particularly relevant for large or geo-distributed federations.
  
  
• CLI Improvements
  CLI commands for creating a collaborator are separated from the certificate request generation, offering improved control of the participant setup workflow.

Highlighting Community Research

Recent efforts to mitigate backdoor attacks in FL include a workflow API example contributed by @perieger (TU Darmstadt) demonstrates how Crowdguard can be used to leverage clients’ feedback on individual models, analyze the behavior of neurons in hidden layers, and eliminate poisoned models through an iterative pruning scheme.

Enhanced Documentation

• Quick Start Tutorial: An updated TaskRunner API Quick Start tutorial makes it easier for newcomers to get up and running with built-in workspaces or custom code.
• FL Plan Description Documentation: There is now a detailed FL plan description in the documentation to help users better understand and configure the federated learning process.

CI/CD

• Ubuntu 22.04 LTS: Docker images have been migrated to the latest Ubuntu 22.04 LTS release for improved stability and support.
• CI Enhancements: Enhanced CI scans and fixes for failing tests ensure a more robust codebase.

Bug Fixes

• FedAvg and NumPy Compatibility: FedAvg in the Workflow interface tutorials is now compatible with the latest NumPy stable release.
• Usability: Various updates including configuration files and fx.init result in a smoother experience.

Get Started with OpenFL 1.6

Find the latest version in the GitHub repository or visit the github.io documentation site for more information. The updated documentation and tutorials are there to guide you every step of the way.

Join the OpenFL Community

The OpenFL community is growing, and we invite you to be a part of it. Join the Slack channel to connect with fellow enthusiasts, share insights, and contribute to the future of federated learning.

Thank you for your continued support. The OpenFL team looks forward to the innovative ways you’ll use OpenFL 1.6 to push the boundaries of federated learning!