Presenting Intel OpenFL 1.6: Federated LLM Assessment and Improvement.
In order to facilitate the upcoming wave of federated learning experiments, the OpenFL team is pleased to announce the release of OpenFL 1.6, which includes a number of new capabilities along with improvements to the documentation and API. For developers working with Large Language Models (LLMs), Intel OpenFL 1.6 is revolutionary because it provides a number of new capabilities that are especially intended to facilitate the shift from a centralized to a federated environment.
By simplifying APIs and offering more instances of OpenFL’s application in cutting-edge situations like federated assessment and federated fine-tuning of LLMs, OpenFL 1.6 aims to enhance the developer experience. This latest version further showcases OpenFL’s expanding compatibility with contemporary computing frameworks and AI accelerators, guaranteeing optimal performance and scalability for federated training and inference.
Along with fixing problems with static code analysis and updating third-party libraries, this version also carries on the security hardening work that serve as the basis for trustworthy federated learning. Code samples and recipes for defending federations against model poisoning attacks are available to assist tackle more complex security issues.
What is OpenFL?
A Python 3 framework for federated learning is called Open Federated Learning (OpenFL). For data scientists, OpenFL is intended to be a versatile, expandable, and simple-to-use platform. The Linux Foundation hosts OpenFL, which strives to be community-driven and encourages contributions.
Why use OpenFL?
With the help of the Python 3 federated learning package OpenFL, businesses may work together to build a model without disclosing private data.
Data sharing concerns are made simpler by FL, but there are still crucial security and privacy factors to take into account. When training AI models in potentially unreliable situations, model creators need to safeguard their intellectual property (IP). By examining model weights across the federated rounds, collaborators must make sure that their data cannot be taken (reverse engineering).
This is where OpenFL comes in; it was created with privacy and security in mind, uses narrow interfaces, and permits all processes to run within Trusted Execution Environments (TEEs) that can guarantee computation integrity, confidentiality of data and models, and enable compute resource attestation.
What’s new With Intel OpenFL 1.6?
With a focus on LLMs, Intel OpenFL 1.6 is jam-packed with changes that address the changing demands of the federated learning community. What to expect from this release is as follows:
Federated LLM Fine-Tuning
Discover how to optimize neuralchat-7b with the OpenFL Workflow API and the Intel Extension for Transformers. Or effectively train LLMs across several private clusters using Horovod.
Workflow API Enhancements
A Workflow API-based FL experiment can be converted into the TaskRunner API format for use in a distributed deployment by utilizing the experimental Workspace Export capability. In addition to the LocalRuntime that is now supported, the foundation is set for a future FederatedRuntime implementation of the Workflow API.
Federated Assessment
By locally verifying the model’s performance on decentralized collaborator nodes and then aggregating these metrics to measure overall effectiveness, federated evaluation enables the assessment of machine learning models in a federated learning system without jeopardizing data confidentiality and privacy. OpenFL now formally supports federated evaluation, and it includes sample tutorials on how to utilize the TaskRunner API to take use of this new feature.
Expanded AI Accelerator Support
Learn how to use XPU (Intel Data Center GPU Max Series) with the Intel Extension for PyTorch. It includes training examples on datasets including Tiny ImageNet (using the Interactive API) and MNIST (using the Workflow API). By using a RayExecutor as a backend for the LocalRuntime of the Workflow API, you may take use of Ray’s GPU management features and multi-node compute scaling. An example notebook for GPU-accelerated model validation on the aggregator is provided to better demonstrate this.
Improved Straggler Collaborator Handling
aggregator fault-tolerance enhancements and bug fixes for situations in which collaborators cease responding or leave a federation. implementing a policy with a cut-off timer and allowing additional rules to be connected. For big or geographically dispersed federations, this capacity is especially essential.
CLI Improvements
Better control over the participant setup process is provided by the separation of CLI commands for collaborator creation from certificate request generation.
Emphasizing Community Research
A workflow API example provided by @perieger (TU Darmstadt) illustrates how Crowdguard can be used to take advantage of client feedback on individual models, examine the behavior of neurons in hidden layers, and remove poisoned models using an iterative pruning scheme. This is one of the recent attempts to mitigate backdoor attacks in FL.
Improved Records
- Quick Start lesson: With the use of custom code or built-in workspaces, a revised TaskRunner API Quick Start lesson makes it simpler for beginners to get started.
- FL Plan Description Documentation: To assist users in comprehending and configuring the federated learning process, a comprehensive FL plan description has been added to the documentation.
CI/CD
- Ubuntu 22.04 LTS: For increased stability and support, Docker images have been updated to the most recent Ubuntu 22.04 LTS version.
- CI Improvements: A more reliable codebase is guaranteed by improved CI scans and corrections for failing tests.
Bug Fixes
- FedAvg and NumPy Compatibility: The most recent stable version of NumPy is now compatible with FedAvg in the Workflow interface tutorials.
- Usability: A smoother experience is the consequence of several modifications, such as those made to fx.init and configuration files.
Conclusion
The federated learning framework for Large Language Models (LLMs), OpenFL 1.6, is announced. With better documentation and APIs, more support for AI accelerators, and enhanced security features, this version aims to improve the developer experience.
New features for federated LLM review and fine-tuning, as well as improved management of straggler contributors, are important enhancements. In order to address security risks like model poisoning, the version also includes community contributions and revised tutorials. Lastly, the announcement promotes involvement and engagement in the community.
