arXiv subjects

Heba Khdr

Publications and source records attributed to Heba Khdr.

3 recordsLinked to original sources

arXiv2025

Accelerated Training on Low-Power Edge Devices

Training on edge devices poses several challenges as these devices are generally resource-constrained, especially in terms of power. State-of-the-art techniques at the device level reduce the GPU frequency to enforce power constraints, leading to a significant increase in training time. To accelerate training, we propose to jointly adjust the system and application parameters (in our case, the GPU frequency and the batch size of the training task) while adhering to the power constraints on devices. We introduce a novel cross-layer methodology that combines predictions of batch size efficiency and device profiling to achieve the desired optimization. Our evaluation on real hardware shows that our method outperforms the current baselines that depend on state of the art techniques, reducing the training time by $2.4\times$ with results very close to optimal. Our measurements also indicate a substantial reduction in the overall energy used for the training process. These gains are achieved without reduction in the performance of the trained model.

cs.LG↗

arXiv2025

Efficient Zeroth-Order Federated Finetuning of Language Models on Resource-Constrained Devices

Federated Learning (FL) is a promising paradigm for finetuning Large Language Models (LLMs) across distributed data sources while preserving data privacy. However, finetuning such large models is challenging on edge devices due to its high resource demand. Zeroth-order Optimization (ZO) estimates gradients through finite-difference approximations, which rely on function evaluations under random perturbations of the model parameters. Consequently, ZO with task alignment provides a potential solution, allowing finetuning using only forward passes with inference-level memory requirements and low communication overhead, but it suffers from slow convergence and higher computational demand. In this paper, we propose a new ZO-based method that applies a more efficient technique to reduce the computational demand associated with using a large number of perturbations while preserving their convergence benefits. This is achieved by splitting the model into consecutive blocks and allocating a higher number of perturbations to the second block, enabling efficient reuse of intermediate activations to update the full network with fewer forward evaluations. Our evaluation on RoBERTa-large, OPT1.3B, LLaMa-3-3.2B models shows up to $3\times$ reduction in computation compared to the other ZO-based techniques, while retaining the memory and communication benefits over first-order federated learning techniques.

cs.LG↗

arXiv2022

NPU-Accelerated Imitation Learning for Thermal Optimization of QoS-Constrained Heterogeneous Multi-Cores

Application migration and dynamic voltage and frequency scaling (DVFS) are indispensable means for fully exploiting the available potential in thermal optimization of a heterogeneous clustered multi-core processor under user-defined quality of service (QoS) targets. However, selecting the core to execute each application and the voltage/frequency (V/f) levels of each cluster is a complex problem because 1) the diverse characteristics and QoS targets of applications require different optimizations, and 2) per-cluster DVFS requires a global optimization considering all running applications. State-of-the-art resource management techniques for power or temperature minimization either rely on measurements that are often not available (such as power) or fail to consider all the dimensions of the problem (e.g., by using simplified analytical models). Imitation learning (IL) enables to use the optimality of an oracle policy, yet at low run-time overhead, by training a model from oracle demonstrations. We are the first to employ IL for temperature minimization under QoS targets. We tackle the complexity by training a neural network (NN) and accelerate the NN inference using a neural processing unit (NPU). While such NN accelerators are becoming increasingly widespread on end devices, they are so far only used to accelerate user applications. In contrast, we use an existing accelerator on a real platform to accelerate NN-based resource management. Our evaluation on a HiKey 970 board with an Arm big.LITTLE CPU and an NPU shows significant temperature reductions at a negligible run-time overhead, with unseen applications and different cooling than used for training.

cs.DC↗