TY - JOUR
AU1 - Deng, Cheng
AU2 - Sun, Luoyang
AU3 - Jiang, Jiwen
AU4 - Zeng, Yongcheng
AU5 - Wu, Xinjian
AU6 - Zhao, Wenxin
AU7 - Xiao, Qingfa
AU8 - Wang, Jiachuan
AU9 - Li, Haoyang
AU1 - Chen, Lei
AU1 - Ni, Lionel M.
AU1 - Zhang, Haifeng
AU1 - Wang, Jun
AB - Abstract:While scaling laws have been continuously validated in large language models (LLMs) with increasing model parameters, the inherent tension between the inference demands of LLMs and the limited resources of edge devices poses a critical challenge to the development of edge intelligence. Recently, numerous small language models have emerged, aiming to distill the capabilities of LLMs into smaller footprints. However, these models often retain the fundamental architectural principles of their larger counterparts, still imposing considerable strain on the storage and bandwidth capacities of edge devices. In this paper, we introduce the PLM, a Peripheral Language Model, developed through a co-design process that jointly optimizes model architecture and edge system constraints. The PLM utilizes a Multi-head Latent Attention mechanism and employs the squared ReLU activation function to encourage sparsity, thereby reducing peak memory footprint during inference. During training, we collect and reorganize open-source datasets, implement a multi-phase training strategy, and empirically investigate the Warmup-Stable-Decay-Constant (WSDC) learning rate scheduler. Additionally, we incorporate Reinforcement Learning from Human Feedback (RLHF) by adopting the ARIES preference learning approach. Following a two-phase SFT process, this method yields performance gains of 2% in general tasks, 9% in the GSM8K task, and 11% in coding tasks. In addition to its novel architecture, evaluation results demonstrate that PLM outperforms existing small language models trained on publicly available data while maintaining the lowest number of activated parameters. Furthermore, deployment across various edge devices, including consumer-grade GPUs, mobile phones, and Raspberry Pis, validates PLM's suitability for peripheral applications. The PLM series models are publicly available at this https URL.
TI - PLM: Efficient Peripheral Language Models Hardware-Co-Designed for Ubiquitous Computing
JF - Computing Research Repository
DO - 10.48550/arxiv.2503.12167
DA - 2025-03-19
UR - https://www.deepdyve.com/lp/arxiv-cornell-university/plm-efficient-peripheral-language-models-hardware-co-designed-for-VHzSJ4QvE3
VL - 2025
IS - 2503
DP - DeepDyve
ER -