A Federated Learning-Inspired Fault-Tolerance Strategy for Large-Scale AI Model Training on HPC Clusters

Develop a fault-tolerance strategy for large-scale AI model training on HPC clusters, drawing inspiration from mechanisms used in federated learning.

The candidate will: 1. Review fault-tolerance techniques for distributed AI training and federated learning. 2. Identify FL mechanisms that may be applicable to HPC training environments. 3. Design and implement a fault-tolerance strategy. 4. Integrate the strategy into a suitable training or experimental framework. 5. Evaluate resilience, overhead, and convergence under representative fault scenarios. 6. Compare the proposed approach with relevant baselines.

6 to 9 months according to the student's background.

• Knowledge of Python
• Knowledge of distributed machine learning
• Knowledge of fault tolerance techniques
• Familiarity with federated learning concepts

1. Gianluca Mittone, & Marco Aldinucci (2026). Accelerating Sharded Data Parallelism at Scale with Federated Learning. In Euro-Par 2026: Parallel Processing - 32nd European Conference on Parallel and Distributed Processing, Pisa, Italy, August 24-28, 2026, Proceedings, Part II (pp. 437–451). Springer.