Edge computing enables distributed inference for computation-intensive applications. However, the autoregressive nature and large model size of large language models (LLMs) pose challenges for their deployment in wireless edge networks. Existing edge inference methods mainly assume an equivalent delay for each token generation step, which fails to capture the dynamic computational and memory overhead incurred during the decoding process. This paper proposes a latency-sensitive wireless edge inference framework for LLMs, where tasks are grouped into multiple batches and processed in parallel by partitioning the LLM into multiple pipeline stages across heterogeneous edge GPUs. An accurate latency model is established, where the latency of each token generation step increases during the autoregressive generation process. Based on this model, the end-to-end inference latency is minimized, which is formulated as a joint optimization problem of bandwidth allocation, model partitioning, and batch scheduling subject to heterogeneous GPU memory constraints. To solve this NP-hard problem with coupled variables, we develop a polynomial-time alternating optimization algorithm that iteratively optimizes model partitioning and batch scheduling via dynamic programming. The closed-form solutions of wireless bandwidth allocation are derived. Extensive simulations show that our approach reduces latency by up to 42.1% versus state-of-the-art baselines across diverse edge scenarios.

The progressing development of Internet of Things (IoT) has accelerated the emergence of resource-intensive and latency-sensitive mobile applications, which throws out a great challenge to the battery-powered wireless devices (WDs) with low-computing capabilities. To solve this intractable issue, we investigate an intelligent reflecting surface (IRS)-assisted multiantenna wireless-powered mobile edge computing (WP-MEC) system, in which WDs first harvest wireless energy emitted by a hybrid access point (HAP), then offload their tasks to the edge server, and finally download the results. In consideration of the practical scenarios, the finite computing capability of edge server and the nonlinear end-to-end power conversion of energy harvesting (EH) circuits at WDs are considered. In addition, an IRS is deployed to improve the efficiency of wireless power transfer (WPT) and the rate of data transmission between HAP and WDs. Under this setup, both space division multiple access (SDMA) and time division multiple access (TDMA) protocols are exploited and evaluated for data transmission. For each protocol, we maximize the computational rate by jointly optimizing time allocation, beamforming designs of HAP and IRS, as well as offloading strategies of WDs. To solve the problem formulated under the SDMA protocol, we propose an efficient alternating optimization (AO) algorithm. For the problem under the TDMA protocol, an AO algorithm with low complexity is proposed. Numerical results demonstrate the high effectiveness of the proposed algorithms and the superiority of the SDMA protocol over the TDMA protocol.

In this paper, we propose an active simultaneously transmitting and reflecting reconfigurable intelligent surface (aSTAR-RIS) assisted wireless information and power transfer system. The aSTAR-RIS is employed to not only assist the downlink simultaneous wireless information and power transfer for the energy harvesting at the wireless powered devices (WPDs) and information receiving at the information receiving devices (IRDs), but also aid the uplink wireless information transmission from the WPDs to the hybrid access point. Our objective is to maximize the uplink sum-rate, considering the quality-of-service requirements of IRDs, by optimizing the power allocation, time scheduling, transmission and reflection beamforming at the aSTAR-RIS. To solve this non-convex problem, we propose an effective alternating optimization algorithm with semidefinite relaxation and fractional programming techniques. Numerical results demonstrate that the proposed aSTAR-RIS scheme can achieve up to 640.68% performance gains compared to the passive STAR-RIS scheme in our settings.