Dongyoon Ryu

This project investigates semantic-aware KV-cache reuse to reduce redundant computation and memory overheads in multi-turn LLM serving. I focus on memory optimization, analyzing KV-cache placement and movement while investigating memory layouts for selective KV reuse and recomputation.

This project develops a scalable framework for distributed deep learning across heterogeneous AI accelerators, including GPUs, NPUs, and PIMs. I am developing an energy profiler to characterize energy consumption across heterogeneous AI accelerators for energy-aware workload partitioning and optimization.

This project develops an ML compiler and runtime framework for heterogeneous on-device AI systems under resource constraints. I am developing a continual learning (CL) system to support efficient on-device model updates while considering latency, memory, and power constraints.