This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.
Permalink
https://hdl.handle.net/2142/124583
Description
Title
A large-scale model serving framework
Author(s)
Zhou, Qinren
Issue Date
2024-04-30
Director of Research (if dissertation) or Advisor (if thesis)
Kindratenko, Volodymyr
Department of Study
Electrical & Computer Eng
Discipline
Electrical & Computer Engr
Degree Granting Institution
University of Illinois at Urbana-Champaign
Degree Name
M.S.
Degree Level
Thesis
Keyword(s)
HPC
LLM
Abstract
Recent advancements in large-scale models have significantly enhanced the capabilities of artificial intelligence. For instance, generative large language models (LLM) have broadly transformed our interactions with websites, devices, and information, in general. These models are increasingly valuable across various domains but pose substantial deployment challenges. They require intense computational resources because of their tens or hundreds of billions of model parameters, necessitating high-end GPUs with large memory capacities. High-performance computing (HPC) clusters, typically equipped with considerable GPU resources, are suitable for serving large-scale models. Nevertheless, their resources are finite and may become inadequate if user requests surpass the cluster’s capacity, unlike the scalable and nearly limitless resources of cloud services. Moreover, HPC clusters lack autoscaling capabilities, introducing additional challenges in accommodating dynamic user demands and optimizing resource utilization.
This thesis presents a model-serving framework to optimize resource utilization and reduce model inference latency. By leveraging the Ray Serve library, the framework automatically manages the deployment and reclamation of models, thereby enabling efficient concurrent service to multiple users. The system incorporates a model-switching mechanism and a Slurm-compatible autoscaler. These features are specifically designed to address traditional HPC clusters’ finite resource and autoscaling limitations and significantly improve system responsiveness and user experience in AI applications.
Use this login method if you
don't
have an
@illinois.edu
email address.
(Oops, I do have one)
IDEALS migrated to a new platform on June 23, 2022. If you created
your account prior to this date, you will have to reset your password
using the forgot-password link below.
