Peng Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Peng Li has a total of 68 co-authors. Here are the top co-authors by number of shared publications:
- Dongxiao Yu
Shandong University
5 shared publications - Deze Zeng
China University of Geosciences
4 shared publications Fahao Chen
School of Artificial Intelligence, Shandong University, Jinan, Shandong, China
4 shared publications- Fahao Chen
Shandong University
4 shared publications - Song Guo
Hong Kong University of Science and Technology, Hong Kong, China
4 shared publications - Xiuzhen Cheng
Shandong University
4 shared publications
Latest Publications
- Open Access
- 22 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 301-310
Queue-Stable Online Learning: Bilinear Bandits for Semantic Codecs in NTNs
- 22 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 301-310
Open AccessQueue-Stable Online Learning: Bilinear Bandits for Semantic Codecs in NTNs
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (11) , 21408-21425
Fair and Efficient Resource Allocation in UAV-Enhanced Edge Computing: A Dual-Layer Auction Approach
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (11) , 21408-21425
Fair and Efficient Resource Allocation in UAV-Enhanced Edge Computing: A Dual-Layer Auction Approach
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (3) , 1963-1974
MAP: GPU Resource Planning for Multi-Agent Systems With Stochastic and Verbose Outputs
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (3) , 1963-1974
MAP: GPU Resource Planning for Multi-Agent Systems With Stochastic and Verbose Outputs
- Open Access
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-25
KVD rive: A Holistic Multi-Tier KV Cache Management System for Long-Context LLM Inference
Supporting long-context LLMs is challenging due to the substantial memory demands of the key–value (KV) cache. Existing offloading systems store the full cache in host memory and selectively fetch critical entries during decoding, but this strategy quickly hits a ceiling: sparsity cannot be pushed further without degrading accuracy. As a result, when context length and batch size grow, the volume of KV transfers rises sharply and becomes the dominant source of decoding latency. We present KVD rive, a holistic multi-tier KV cache management system spanning GPU memory, host DRAM, and SSD. Unlike prior work that pursues greater sparsity through algorithmic refinements, KVD rive tackles the problem from a systems perspective—jointly orchestrating cache placement, pipeline scheduling, and cross-tier coordination to sustain high-throughput inference under tight GPU budgets. KVD rive advances three fundamental capabilities: it adapts cache management to attention behavior to maximize reuse and minimize redundant data movement; it restructures the decoding pipeline to overlap I/O- and CPU/GPU compute-bound stages, eliminating stalls across heterogeneous resources; and it harmonizes data movement across memory tiers to unlock scalable long-context inference far beyond GPU and DRAM limits. We have implemented a fully functional prototype of KVD rive and evaluated it on long-context benchmarks with popular LLMs. The system achieves up to 1.74× higher throughput compared to state-of-the-art works while preserving accuracy.
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-25
Open AccessKVD rive: A Holistic Multi-Tier KV Cache Management System for Long-Context LLM Inference
Supporting long-context LLMs is challenging due to the substantial memory demands of the key–value (KV) cache. Existing offloading systems store the full cache in host memory and selectively fetch critical entries during decoding, but this strategy quickly hits a ceiling: sparsity cannot be pushed further without degrading accuracy. As a result, when context length and batch size grow, the volume of KV transfers rises sharply and becomes the dominant source of decoding latency. We present KVD rive, a holistic multi-tier KV cache management system spanning GPU memory, host DRAM, and SSD. Unlike prior work that pursues greater sparsity through algorithmic refinements, KVD rive tackles the problem from a systems perspective—jointly orchestrating cache placement, pipeline scheduling, and cross-tier coordination to sustain high-throughput inference under tight GPU budgets. KVD rive advances three fundamental capabilities: it adapts cache management to attention behavior to maximize reuse and minimize redundant data movement; it restructures the decoding pipeline to overlap I/O- and CPU/GPU compute-bound stages, eliminating stalls across heterogeneous resources; and it harmonizes data movement across memory tiers to unlock scalable long-context inference far beyond GPU and DRAM limits. We have implemented a fully functional prototype of KVD rive and evaluated it on long-context benchmarks with popular LLMs. The system achieves up to 1.74× higher throughput compared to state-of-the-art works while preserving accuracy.
- 1 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (5) , 1048-1061
Fed-Grow: Federating to Grow Transformers for Resource-Constrained Users Without Model Sharing
The growing resource demands of large-scale transformer models pose significant challenges for resource-constrained users, particularly in distributed environments. To address this issue, we propose a federated learning framework called Fed-Grow, which enables multiple participants to collaboratively learn a lightweight scaling operation that transfers knowledge from pretrained small models to a large transformer model. In Fed-Grow, we introduce the Dual-LiGO (Dual Linear Growth Operator) architecture, consisting of Local-LiGO and Global-LiGO components. Local-LiGO addresses model heterogeneity by adapting each participant’s pre-trained model to a common intermediate form, while Global-LiGO facilitates knowledge sharing across participants without sharing local models or raw data, ensuring privacy preservation. This federated approach offers a scalable solution for growing large transformers in a distributed manner, where only the Global-LiGO is shared, significantly reducing communication overhead while maintaining comparable model performance under the same communication constraints. Experimental results demonstrate that Fed-Grow outperforms state-of-the-art methods in terms of accuracy and precision, while reducing the number of trainable parameters by 59.25% and communication costs by 73.01% . These improvements allow for higher efficiency in training large models in distributed environments, without sacrificing performance. To the best of our knowledge, Fed-Grow is the first method that enables cooperative transformer scaling in a distributed setting, making it a practical solution for resource-constrained users.
- 1 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 37 (5) , 1048-1061
Fed-Grow: Federating to Grow Transformers for Resource-Constrained Users Without Model Sharing
The growing resource demands of large-scale transformer models pose significant challenges for resource-constrained users, particularly in distributed environments. To address this issue, we propose a federated learning framework called Fed-Grow, which enables multiple participants to collaboratively learn a lightweight scaling operation that transfers knowledge from pretrained small models to a large transformer model. In Fed-Grow, we introduce the Dual-LiGO (Dual Linear Growth Operator) architecture, consisting of Local-LiGO and Global-LiGO components. Local-LiGO addresses model heterogeneity by adapting each participant’s pre-trained model to a common intermediate form, while Global-LiGO facilitates knowledge sharing across participants without sharing local models or raw data, ensuring privacy preservation. This federated approach offers a scalable solution for growing large transformers in a distributed manner, where only the Global-LiGO is shared, significantly reducing communication overhead while maintaining comparable model performance under the same communication constraints. Experimental results demonstrate that Fed-Grow outperforms state-of-the-art methods in terms of accuracy and precision, while reducing the number of trainable parameters by 59.25% and communication costs by 73.01% . These improvements allow for higher efficiency in training large models in distributed environments, without sacrificing performance. To the best of our knowledge, Fed-Grow is the first method that enables cooperative transformer scaling in a distributed setting, making it a practical solution for resource-constrained users.
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2292-2307
Efficient Multimodal Serving via Module Multiplexing
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2292-2307
Open AccessEfficient Multimodal Serving via Module Multiplexing
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 548-563
Federated Fine-Tuning of Sparsely-Activated Large Language Models on Resource-Constrained Devices
Federated fine-tuning of Mixture-of-Experts (MoE)-based large language models (LLMs) is challenging due to their massive computational requirements and the resource constraints of participants. Existing works attempt to fill this gap through model quantization, computation offloading, or expert pruning. However, they cannot achieve desired performance due to impractical system assumptions and a lack of consideration for MoE-specific characteristics. In this paper, we propose Flux, a system designed to enable federated fine-tuning of MoE-based LLMs across participants with constrained computing resources (e.g., consumer-grade GPUs), aiming to minimize time-to-accuracy. Flux introduces three key innovations: (1) quantization-based local profiling to estimate expert activation with minimal overhead, (2) adaptive layer-aware expert merging to reduce resource consumption while preserving accuracy, and (3) dynamic expert role assignment using an exploration-exploitation strategy to balance tuning and non-tuning experts. Extensive experiments on LLaMA-MoE and DeepSeek-MoE with multiple benchmark datasets demonstrate that Flux significantly outperforms existing methods, achieving up to 4.75× speedup in time-to-accuracy.
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 548-563
Open AccessFederated Fine-Tuning of Sparsely-Activated Large Language Models on Resource-Constrained Devices
Federated fine-tuning of Mixture-of-Experts (MoE)-based large language models (LLMs) is challenging due to their massive computational requirements and the resource constraints of participants. Existing works attempt to fill this gap through model quantization, computation offloading, or expert pruning. However, they cannot achieve desired performance due to impractical system assumptions and a lack of consideration for MoE-specific characteristics. In this paper, we propose Flux, a system designed to enable federated fine-tuning of MoE-based LLMs across participants with constrained computing resources (e.g., consumer-grade GPUs), aiming to minimize time-to-accuracy. Flux introduces three key innovations: (1) quantization-based local profiling to estimate expert activation with minimal overhead, (2) adaptive layer-aware expert merging to reduce resource consumption while preserving accuracy, and (3) dynamic expert role assignment using an exploration-exploitation strategy to balance tuning and non-tuning experts. Extensive experiments on LLaMA-MoE and DeepSeek-MoE with multiple benchmark datasets demonstrate that Flux significantly outperforms existing methods, achieving up to 4.75× speedup in time-to-accuracy.
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (2) , 579-592
DAHFF: Joint Device Selection and Bandwidth Allocation for Efficient Hierarchical Federated Learning
Federated learning, as a compelling machine learning framework, enables collaborative model training without exposing private data. However, the excessive communication overhead remains a major challenge. To tackle this challenge, hierarchical federated edge learning (HFEL) framework has been proposed for reducing the communication load via migrating the model aggregation partially from cloud to edge servers. Although HFEL has significant potential, it is still constrained by end-devices with limited computational capabilities and unfavorable network conditions. A common approach to reduce this effect is to involve only the fastest end-devices in the training process. But because only parts of end-devices' data samples can be selected by such means, it damages the diversity of training data, and hence greatly affects the model's quality. In addition, for further improving the training performance, a proper bandwidth allocation strategy is also needed to make full use of the shared network resource of edge servers. To this end, we propose DAHFF, a Diversity-Aware Hierarchical Fast Federated learning framework consisting of Virtual Queue based Device Selection phase and Binary Search based Bandwidth Allocation, which are responsible for selecting participated end-devices and allocating bandwidth for selected devices, respectively. Extensive experiments on different deep learning models show that our proposed framework can averagely speed up the training performance by $2.07\times$ in comparison with state-of-the-art approaches.
- 23 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (2) , 579-592
DAHFF: Joint Device Selection and Bandwidth Allocation for Efficient Hierarchical Federated Learning
Federated learning, as a compelling machine learning framework, enables collaborative model training without exposing private data. However, the excessive communication overhead remains a major challenge. To tackle this challenge, hierarchical federated edge learning (HFEL) framework has been proposed for reducing the communication load via migrating the model aggregation partially from cloud to edge servers. Although HFEL has significant potential, it is still constrained by end-devices with limited computational capabilities and unfavorable network conditions. A common approach to reduce this effect is to involve only the fastest end-devices in the training process. But because only parts of end-devices' data samples can be selected by such means, it damages the diversity of training data, and hence greatly affects the model's quality. In addition, for further improving the training performance, a proper bandwidth allocation strategy is also needed to make full use of the shared network resource of edge servers. To this end, we propose DAHFF, a Diversity-Aware Hierarchical Fast Federated learning framework consisting of Virtual Queue based Device Selection phase and Binary Search based Bandwidth Allocation, which are responsible for selecting participated end-devices and allocating bandwidth for selected devices, respectively. Extensive experiments on different deep learning models show that our proposed framework can averagely speed up the training performance by $2.07\times$ in comparison with state-of-the-art approaches.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 5942-5953
Efficient Mixture-of-Experts Model Inference at the Edge via Adaptive Expert Merging
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Networking
- Vol. 34, 5942-5953
Efficient Mixture-of-Experts Model Inference at the Edge via Adaptive Expert Merging
- 30 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (6) , 8830-8843
Fed-RAA: Resource-Adaptive Asynchronous Federated Edge Learning With Theoretical Guarantee
This paper studies an efficient federated learning (FL) problem involving multiple edge-based clients with heterogeneous constrained resources. Compared with numerous training parameters, the computing and communication resources of clients in edge scenarios are usually insufficient for fast local training and real-time knowledge sharing. Besides, training on clients with heterogeneous resources may result in the straggler problem, which delays the convergence of FL. To address these issues, we propose Fed-RAA: a Resource-Adaptive Asynchronous Federated learning algorithm. Different from vanilla FL methods, where all parameters are trained by each participating client regardless of resource diversity, Fed-RAA adaptively allocates submodels of the global model to clients based on their computing and communication capabilities. Each client then individually trains its assigned submodel and asynchronously uploads the updated result. Theoretical analysis confirms the convergence of our approach. Additionally, an online greedy-based algorithm is designed for asynchronous submodel assignment in Fed-RAA, improving the convergence of Fed-RAA by optimal minimization on the training delay bound of submodels. Compared to state-of-the-art methods, our Fed-RAA algorithm reduces the time required to achieve the target accuracy by an average of$30.89\%$ , demonstrating its superior efficiency on heterogeneous constrained computing and communication resources. To the best of our knowledge, this paper is the first resource-adaptive asynchronous method for submodel-based FL with guaranteed theoretical convergence.- 30 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (6) , 8830-8843
Fed-RAA: Resource-Adaptive Asynchronous Federated Edge Learning With Theoretical Guarantee
This paper studies an efficient federated learning (FL) problem involving multiple edge-based clients with heterogeneous constrained resources. Compared with numerous training parameters, the computing and communication resources of clients in edge scenarios are usually insufficient for fast local training and real-time knowledge sharing. Besides, training on clients with heterogeneous resources may result in the straggler problem, which delays the convergence of FL. To address these issues, we propose Fed-RAA: a Resource-Adaptive Asynchronous Federated learning algorithm. Different from vanilla FL methods, where all parameters are trained by each participating client regardless of resource diversity, Fed-RAA adaptively allocates submodels of the global model to clients based on their computing and communication capabilities. Each client then individually trains its assigned submodel and asynchronously uploads the updated result. Theoretical analysis confirms the convergence of our approach. Additionally, an online greedy-based algorithm is designed for asynchronous submodel assignment in Fed-RAA, improving the convergence of Fed-RAA by optimal minimization on the training delay bound of submodels. Compared to state-of-the-art methods, our Fed-RAA algorithm reduces the time required to achieve the target accuracy by an average of$30.89\%$ , demonstrating its superior efficiency on heterogeneous constrained computing and communication resources. To the best of our knowledge, this paper is the first resource-adaptive asynchronous method for submodel-based FL with guaranteed theoretical convergence.