Efficient Generative AI

Efficiency is a fundamental enabler for scaling generative AI to real-world deployment. Our research develops principled methods to improve the performance, cost-efficiency, and scalability of large models across modalities. We focus on Transformer and LLM optimization—pioneering techniques such as SpAtten (cascade KV cache pruning and quantization), SpAtten-Chip, Hardware-Aware Transformer, Lightning-Transformer, and SpArch for system-algorithm co-design. In computer vision, we introduced AMC, and APQ for automated model compression. Our work leverages pruning, quantization, neural architecture search, reinforcement learning, and compiler-hardware-algorithm co-design to build highly optimized models that run efficiently on diverse hardware—from cloud GPUs to edge devices. These innovations have been widely adopted in both academia and industry, powering faster, more efficient, and accessible foundation models.

Efficient AI Systems with Emerging Technology

We explore how emerging compute platforms—such as quantum computers and photonics accelerators—can advance AI efficiency and scalability. Our work spans AI-centric hardware-software co-design, photonic neural networks (Lightning-Transformer), hybrid quantum-classical AI acceleration (QuantumNAS, QuEst, QuantumNAT, QOC, Atomique), and system-level optimization to push the limits of model training and inference performance.