J. Pablo Muñoz

Sixing Yu, J. Pablo Muñoz, Ali Jannesari

Federated learning (FL) offers privacy-preserving decentralized machine learning, optimizing models at edge clients without sharing private data. Simultaneously, foundation models (FMs) have gained traction in the artificial intelligence (AI) community due to their exceptional performance across various tasks. However, integrating FMs into FL presents challenges, primarily due to their substantial size and intensive resource requirements. This is especially true when considering the resource heterogeneity in edge FL systems. We present an adaptive framework for Resource-aware Federated Foundation Models (RaFFM) to address these challenges. RaFFM introduces specialized model compression algorithms tailored for FL scenarios, such as salient parameter prioritization and high-performance subnetwork extraction. These algorithms enable dynamic scaling of given transformer-based FMs to fit heterogeneous resource constraints at the network edge during both FL's optimization and deployment stages. Experimental results demonstrate that RaFFM shows significant superiority in resource utilization efficiency and uses fewer resources to deploy FMs to FL. Despite the lower resource consumption, target models optimized by RaFFM achieve performance on par with traditional FL methods applied to full-sized FMs. This is evident across tasks in both natural language processing and computer vision domains.

Md Anwar Hossen, Fatema Siddika, Juan Pablo Munoz et al.

Large language models (LLMs) can acquire new capabilities through fine-tuning, but continual adaptation often leads to catastrophic forgetting. We propose CRAFT, a continual learning framework that avoids updating model weights by instead learning low-rank interventions on hidden representations. CRAFT proceeds in three stages: it first routes each task to a group of similar tasks based on output-distribution divergence; it then fine-tunes the model using a Kullback-Leibler (KL) divergence against the group's prior state, which directly controls forgetting and determines convergence; finally, it merges interventions for the updated task into the shared representation using the same KL signal. This design unifies routing, regularization, and merging through a single KL-based objective. CRAFT improves overall performance and reduces forgetting compared to strong LoRA-based approaches across multiple benchmarks and model scales, while remaining robust to task ordering. These results suggest that controlling adaptation in representation space, guided by output-space divergence, provides a scalable and principled approach to continual learning in LLMs.

Abdullah Al Asif, Patrick Diem, Juan Pablo Muñoz et al.

Distributed Low-Communication (DiLoCo) training reduces communication overhead by allowing workers to perform multiple local optimization steps before sending pseudo-gradients to a global outer update. Its asynchronous variant further improves hardware utilization by removing synchronization barriers, but at the cost of stale pseudo-gradients computed from outdated model states. As a result, these updates can become misaligned with the current global optimization direction, particularly in heterogeneous systems. This issue becomes even more pronounced when data are non-IID, a setting that has not been well studied in asynchronous low-communication training. To address this limitation, we propose \textbf{HeLoCo}, a direction-aware correction method for asynchronous low-communication training that uses outer momentum as a reference for the current optimization trajectory and selectively adjusts incoming pseudo-gradients before the outer update. Updates that remain aligned are preserved, while directionally conflicting components are corrected. On multilingual language-model training with heterogeneous workers and non-IID data, HeLoCo consistently improves validation loss. It outperforms existing asynchronous DiLoCo-based baselines by up to 7.5\% at a fixed token budget, exceeds asynchronous momentum look-ahead by up to 3.3\% at a fixed wall-clock budget, and surpasses the synchronous baseline by up to 22.1\% under severe system heterogeneity. Our analysis further shows how staleness, worker speed, and data heterogeneity shape update quality and convergence in highly decentralized and heterogeneous training setups.

J. Pablo Muñoz, Jinjie Yuan, Nilesh Jain

Large pre-trained models have achieved outstanding results in sequence modeling. The Transformer block and its attention mechanism have been the main drivers of the success of these models. Recently, alternative architectures, such as Selective Structured State Space Models (SSMs), have been proposed to address the inefficiencies of Transformers. This paper explores the compression of SSM-based models, particularly Mamba and its hybrids. We study the sensitivity of these models to the removal of selected components at different granularities to reduce the model size and computational overhead, thus improving their efficiency while maintaining accuracy. The proposed solutions, collectively referred to as Mamba-Shedder, achieve a speedup of up to 1.4x during inference, demonstrating that model efficiency can be improved by eliminating several redundancies with minimal impact on the overall model performance. The code is available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.

Abdullah Al Asif, Sixing Yu, Juan Pablo Munoz et al.

SplitFed Learning (SFL) combines federated learning and split learning to enable collaborative training across distributed edge devices; however, it faces significant challenges in heterogeneous environments with diverse computational and communication capabilities. This paper proposes \textit{SuperSFL}, a federated split learning framework that leverages a weight-sharing super-network to dynamically generate resource-aware client-specific subnetworks, effectively mitigating device heterogeneity. SuperSFL introduces Three-Phase Gradient Fusion (TPGF), an optimization mechanism that coordinates local updates, server-side computation, and gradient fusion to accelerate convergence. In addition, a fault-tolerant client-side classifier and collaborative client--server aggregation enable uninterrupted training under intermittent communication failures. Experimental results on CIFAR-10 and CIFAR-100 with up to 100 heterogeneous clients show that SuperSFL converges $2$--$5\times$ faster in terms of communication rounds than baseline SFL while achieving higher accuracy, resulting in up to $20\times$ lower total communication cost and $13\times$ shorter training time. SuperSFL also demonstrates improved energy efficiency compared to baseline methods, making it a practical solution for federated learning in heterogeneous edge environments.

Ahmed F. AbouElhamayed, Jordan Dotzel, Yash Akhauri et al.

Large language models have high compute, latency, and memory requirements. While specialized accelerators such as GPUs and TPUs typically run these workloads, CPUs are more widely available and consume less energy. Accelerating LLMs with CPUs enables broader AI access at a lower cost and power consumption. This acceleration potential for CPUs is especially relevant during the memory-bound decoding stage of LLM inference, which processes one token at a time and is becoming increasingly utilized with reasoning models. We utilize Advanced Matrix Extensions (AMX) support on the latest Intel CPUs together with unstructured sparsity to achieve a $1.42 \times$ reduction in end-to-end latency compared to the current PyTorch implementation by applying our technique in linear layers. We provide a set of open-source customized sparse kernels that can speed up any PyTorch model by automatically replacing all linear layers with our custom sparse implementation. Furthermore, we demonstrate for the first time the use of unstructured sparsity in the attention computation achieving a $1.14 \times$ speedup over the current systems without compromising accuracy. Code: https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning/tree/main/SparAMX

J. Pablo Muñoz, Jinjie Yuan, Nilesh Jain

The rapid expansion of Large Language Models (LLMs) has posed significant challenges regarding the computational resources required for fine-tuning and deployment. Recent advancements in low-rank adapters have demonstrated their efficacy in parameter-efficient fine-tuning (PEFT) of these models. This retrospective paper comprehensively discusses innovative approaches that synergize low-rank representations with Neural Architecture Search (NAS) techniques, particularly weight-sharing super-networks. Robust solutions for compressing and fine-tuning large pre-trained models are developed by integrating these methodologies. Our analysis highlights the potential of these combined strategies to democratize the use of LLMs, making them more accessible for deployment in resource-constrained environments. The resulting models exhibit reduced memory footprints and faster inference times, paving the way for more practical and scalable applications of LLMs. Models and code are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.

J. Pablo Muñoz, Jinjie Yuan, Nilesh Jain

Recently, state-of-the-art approaches for pruning large pre-trained models (LPMs) have demonstrated that the training-free removal of non-critical residual blocks in Transformers is viable for reducing model size, achieving results that outperform previous training-free pruning approaches. Motivated by these findings, we extend BlockPruner (Zhong et al., 2024) and propose MultiPruner, a pruning approach that surpasses recent training-free pruning methods by adopting a multidimensional, iterative, fine-grained pruning strategy. In MultiPruner, multidimensional pruning reinstates the structural balance in block-pruned models by sequentially compressing along three dimensions: i) residual blocks, ii) channels of multilayer perceptrons (MLP), and iii) attention heads. This solution enhances zero-shot accuracy on downstream tasks compared to other techniques while improving model compression ratios, producing compressed models with fewer computing and memory requirements. Extensive experiments demonstrate the advantages of the proposed method across various large pre-trained models. The code and pruning configurations are available at https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.

Sixing Yu, J. Pablo Muñoz, Ali Jannesari

Foundation Models (FMs), such as LLaMA, BERT, GPT, ViT, and CLIP, have demonstrated remarkable success in a wide range of applications, driven by their ability to leverage vast amounts of data for pre-training. However, optimizing FMs often requires access to sensitive data, raising privacy concerns and limiting their applicability in many domains. In this paper, we propose the Federated Foundation Models (FFMs) paradigm, which combines the benefits of FMs and Federated Learning (FL) to enable privacy-preserving and collaborative learning across multiple end-users. We discuss the potential benefits and challenges of integrating FL into the lifespan of FMs, covering pre-training, fine-tuning, and application. We further outline potential future research avenues in FFM, including FFM pre-training, FFM fine-tuning, and federated prompt tuning, which allow the development of more personalized and context-aware models while ensuring data privacy. Moreover, we explore the possibility of continual/lifelong learning in FFMs, as increased computational power at the edge may unlock the potential for optimizing FMs using newly generated private data close to the data source. The proposed FFM concepts offer a flexible and scalable framework for training large language models in a privacy-preserving manner, setting the stage for subsequent advancements in both FM training and federated learning.

J. Pablo Muñoz, Nikolay Lyalyushkin, Yash Akhauri et al.

Recent Neural Architecture Search (NAS) solutions have produced impressive results training super-networks and then deriving subnetworks, a.k.a. child models that outperform expert-crafted models from a pre-defined search space. Efficient and robust subnetworks can be selected for resource-constrained edge devices, allowing them to perform well in the wild. However, constructing super-networks for arbitrary architectures is still a challenge that often prevents the adoption of these approaches. To address this challenge, we present BootstrapNAS, a software framework for automatic generation of super-networks for NAS. BootstrapNAS takes a pre-trained model from a popular architecture, e.g., ResNet- 50, or from a valid custom design, and automatically creates a super-network out of it, then uses state-of-the-art NAS techniques to train the super-network, resulting in subnetworks that significantly outperform the given pre-trained model. We demonstrate the solution by generating super-networks from arbitrary model repositories and make available the resulting super-networks for reproducibility of the results.