전명재 연구실

System-level optimization for Large Language Model serving that dynamically balances memory usage and performance to meet specific Service Level Objectives (SLOs). Includes SLO-aware adaptive KV cache management, generalized KV cache reuse, accuracy-preserving long-context pruning, failure recovery, and systems for agentic AI.

AI paradigm enabling local devices to train and adapt models instantly without cloud servers. Focuses on low-latency, high-accuracy on-device learning and adaptive, resource-efficient continual learning from constant streams of new data.

Intelligent orchestration strategy for training large AI models across multiple, physically separated data centers with varying hardware capabilities. Includes learning-based planning for heterogeneous geo-distributed training and fast distributed training on heterogeneous accelerators.

Strategy for speeding up data-heavy tasks by intelligently storing and reusing data. Covers efficient caching for iterative analytics and data preprocessing for scalable ML pipelines handling millions or billions of records.