Daliang Xu

Daliang Xu

Assistant Professor (Associate Researcher) Beijing University of Posts and Telecommunications

Daliang Xu （徐大亮）

I am an incoming Assistant Professor (Associate Researcher) at Beijing University of Posts and Telecommunications (BUPT), and I will soon work with Prof. Shangguang Wang and Prof. Mengwei Xu. I received my Ph.D. from Peking University (PKU) in June 2025, where I was fortunate to be advised by Prof. Gang Huang, Prof. Xuanzhe Liu, and Prof. Mengwei Xu. My research interests are in mobile computing and system software.

I always look for highly self-motivated undergraduates and graduates. If you are interested in my research, please feel free to send your CV to contact me at bupt_on_device_lab@163.com.

Notice: Please refer to the tips before sending your email.

Research

My research focuses on empowering resource-constrained edge devices (e.g., satellites, UAVs, and smartphones) with multimodal large language model (LLM) capabilities through hardware-software co-design.
  • Efficient on-device multimodal LLMs
    - Our research primarily optimizes on-device multimodal LLM inference from the perspective of heterogeneous hardware.
    • Heterogeneous computing systems (e.g., NPU) for on-device multimodal LLMs.
      - Mobile devices typically contain a variety of heterogeneous computing resources (such as CPU, GPU, NPU, etc.). However, current on-device multimodal LLM systems fail to fully utilize them. To address this, we are designing new system software stack optimized for heterogeneous computing resources to maximize their utilization.

      Our current research topics cover NPU-optimized on-device multimodal LLM engines, NPU compiler design, and other related areas.

      Papers: Mandheling [MobiCom22], Niagara [ICSOC23 Distinguished Award], SoCFlow [ASPLOS24], EdgeLLM [TMC24], LLM.NPU() [ASPLOS25]
    • NPU-friendly multimodal LLM algorithms.
      PieBridge [SenSys24], Q-FedUpdate [WWW24]
    • Intelligent autonomous systems: multimodal perception, autonomous control, and self-evolution in UAVs or satellites.
  • New hardware for intelligent satellites or smartphones.
    - Next-generation intelligent satellites featuring high reliability, fault tolerance, and support for multimodal LLM inference.
    • On-device accelerators for multimodal LLM.
      - Our research focuses on multimodal LLM quantized inference efficiency and minimizing accelerator energy consumption, current, and area.
    • High-Reliability SoCs for Satellites.
      - Focus on characteristics such as high reliability and fault tolerance.

On-going projects

  • MLLM-NPU - a fast and lightweight NPU-optimized multimodal LLM inference engine for mobile devices.

  • Satellite hardware. - A lightweight satellite SoC, focusing on space-grade reliability and enhancing multimodal LLM acceleration capabilities.

Selected Publications (* = equal contributions)

• [CCF-A] [ASPLOS'2025] Fast On-device LLM Inference with NPUs
Daliang Xu, Hao Zhang, Liming Yang, Ruiqi Liu, Gang Huang, Mengwei Xu, Xuanzhe Li
• [CCF-A] [ASPLOS'2024] SoCFlow: Efficient and Scalable DNN Training on SoC-Clustered Edge Servers
Daliang Xu*, Mengwei Xu*, Chiheng Lou, Li Zhang, Gang Huang, Xin Jin, Xuanzhe Liu
• [CCF-A] [TMC'2024] EdgeLLM: Fast On-Device LLM Inference With Speculative Decoding
Daliang Xu, Wangsong Yin, Hao Zhang, Xin Jin, Ying Zhang, Shiyun Wei, Mengwei Xu, Xuanzhe Liu
• 🏆 [CCF-B Distinguished Paper Award] [ICSOC'2023] Niagara: Scheduling DNN Inference Services on Heterogeneous Edge Processors
Daliang Xu, Qing Li, Mengwei Xu, Kang Huang, Gang Huang, Shangguang Wang, Xin Jin, Ma Yun, Xuanzhe Liu
• [CCF-A] [Mobicom'2022] Mandheling: Mixed-Precision On-Device DNN Training with DSP Offloading
Daliang Xu*, Mengwei Xu*, Qipeng Wang, Shangguang Wang, Kang Huang, Gang Huang, Xin Jin, Xuanzhe Liu