Suraiya Tairin, Ph.D.

Improving the efficiency, scalability, and reliability of LLM inference by optimizing multiple system-level factors, including resource, cluster, scheduling, communication, and deployment efficiency.

Relevant works

• eMoE: Task-aware Memory Efficient Mixture-of-Experts-Based (MoE) Model Inference (Preprint ↗)

Accelerating distributed deep learning training on multi-GPU clusters by detecting and mitigating slow workers (stragglers).

Relevant works

• Towards Efficient Straggler Management in Distributed Deep Learning Training
• Revisiting the Straggling Problem in GPU-based Distributed Deep Learning Training
• Proactive, Accuracy-aware Straggler Mitigation in Machine Learning Clusters

Resource-aware scheduling that allocates GPU resources to ML training jobs efficiently and fairly as resource availability changes.

Relevant works

• Revisiting the Straggling Problem in GPU-based Distributed Deep Learning Training
• Embracing Uncertainty for Equity in Resource Allocation in ML Training

LLM-based agents that plan, reason, use tools, and coordinate multi-step workflows, with a focus on making them efficient and reliable.

Efficient training and serving of generative AI models, including LLMs and vision-language models, on GPUs.

Relevant works

• eMoE: Task-aware Memory Efficient Mixture-of-Experts-Based (MoE) Model Inference (Preprint ↗)

Computer vision, multimodal AI, and data science for real-world problems, such as pavement distress and rail fault detection in transportation.