Research
The problems we work on
Cloud-Edge Collaborative LLM Serving
Edge devices have limited compute, while cloud models are powerful but distant and costly. We study how to split requests, caches, and computation between cloud and edge to improve the latency, throughput, and quality of LLM services.
End-to-End Acceleration of LLM Applications
Real-world LLM applications are workflows of RAG, agents, and multiple models. We optimize entire workflows end to end on heterogeneous hardware.
Efficient Generative AI
Diffusion models for image and video generation run dozens of denoising steps. We exploit redundancy across steps with feature caching, feature forecasting, and multi-device collaboration to accelerate generation with little quality loss.
Edge Container Scheduling and Resource Management
Edge clusters are resource-constrained and highly dynamic. We study layer-aware container scheduling, migration, and updates, cold-start optimization, and resource overbooking, using deep reinforcement learning and online algorithms.
Efficient Multimodal Inference and Retrieval
Retrieval and question answering over images, text, audio, and video must be accurate, fast, and memory-efficient on edge devices.
Task Scheduling and Optimization in Edge Computing
Our long-standing work covers task offloading and scheduling, service placement, UAV and vehicular edge computing, crowdsourcing, and federated learning, as well as a survey of on-device AI models.