Artifact of the paper: Accelerating Sharded Data Parallelism at Scale with Federated Learning

This artifact accompanies the EuroPar paper \emph{“Accelerating Sharded Data Parallelism at Scale with Federated Learning”}. It provides implementations of the proposed hybrid training algorithms (FL+FSDP and FL+HSDP), along with their standard counterparts (FSDP and HSDP).

The artifact enables reviewers to reproduce the experiments that demonstrate the correctness of the hybrid FL + sharded data-parallel workflow and the communication and computation behaviour, as well as the representative results and plots from the paper.

Detailed reproducibility instructions are given in the examples/EuroPar/README.md file. Additional details about xFFL are available in its repository branch dedicated to this artifact evaluation: https://github.com/alpha-unito/xffl/tree/FL+DP/examples/EuroPar.