DB for AI Group

Overview

Efficient AI training and inference

DB for AI develops data management techniques for efficient AI training and inference. The group studies the full processing pipeline, from data preparation and model compilation to batch generation, data transfer, approximate computation and workload scheduling. Its research includes graph neural network training and large language model serving.

The aim is to reduce energy consumption and development and deployment costs, shorten AI development cycles, and improve processing speed.

Research areas

• Data preparation. Coreset selection and graph coarsening for training workloads. [Camel, CC-GNN]
• Model compilation. Graph rewriting, sparse formats and efficient tensor program reuse. [STile, ETO]
• Batch generation. Feature-oriented sampling and dynamic data placement for graph training. [SIMPLE, ETC]
• Data transfer. Locality-aware transfer and decentralized training with cached data. [DUCATI, SANCUS]
• Approximate computation. Historical caching, approximate query processing and knowledge distillation. [Orca, Zebra, Graph distillation]
• Workload scheduling. Hybrid caching and request scheduling for scalable LLM inference. [Apt-Serve]