Collaborative Research: Frameworks: hpcGPT: Enhancing Computing Center User Support with HPC-enriched Generative AI

hpcGPT是为学术计算中心（如国家超级计算应用中心、俄亥俄超级计算机中心、圣地亚哥超级计算机中心和德克萨斯高级计算中心）提供的问答服务。这些中心为成千上万的用户提供高性能计算（HPC）平台，用于科学和工程研究。hpcGPT与美国普林斯顿大学和罗格斯大学合作，采用生成式人工智能技术，整合不同更新频率的异构数据源，提高用户支持服务质量和效率，缩短响应时间，提高支持精度。使用hpcGPT，用户支持团队可以利用历史知识、实时系统状态和外部技术专长来更好地支持HPC用户。通过hpcGPT高质量、及时的解答，HPC用户可以解决许多技术问题，从而减少用户支持团队的工作量。这将允许支持团队更多地关注新的和新颖的支持问题。hpcGPT将显著提高用户支持服务的质量、容量和效率，而无需增加人力。hpcGPT结合了微调和检索增强生成（Retrieval Augmented Generation， RAG）技术，结合了通用计算的最新知识、过去的经验、领域专业知识、文档和实时系统状态。hpcGPT基于大型语言模型微调和托管、检索增强生成和外部数据源集成等现有和公认的功能，降低了调整信息和识别问题、答案和支持信息之间的依赖关系所需的复杂性和工作量。这对于具有不同应用程序需求和有限人员的研究小组和计算中心尤其有益。hpcGPT扩展和翻译了一套网络基础设施构建块和技术，如大型语言模型训练和推理服务托管。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

hpcGPT is a question answering service for academic computing centers such as the National Center for Supercomputing Applications, Ohio Supercomputer Center, San Diego Supercomputer Center, and Texas Advanced Computing Center. These Centers provide high-performance computing (HPC) platforms to tens of thousands of users for science and engineering research. In collaboration with Princeton University and Rutgers University, hpcGPT uses generative artificial intelligence (AI) and integrates heterogeneous data sources with different update frequencies to enhance the user support service quality and efficiency, decrease the response time, and improve precision of the support. With hpcGPT, user support teams can leverage the historical knowledge, real-time system status, and external technical expertise to better support the HPC users. With the high-quality and timely answers from hpcGPT, HPC users can resolve many technical issues, thus reducing the workload of the user support teams. This will allow the support teams to focus more on new and novel support issues. hpcGPT will significantly enhance the user support service quality, capacity, and efficiency without increasing the human effort.hpcGPT combines the fine-tuning and Retrieval Augmented Generation (RAG) techniques to incorporate recent knowledge, past experience, domain expertise, documentations, and real-time system status of versatile computing. By building upon existing and recognized capabilities in large language model fine-tuning and hosting, retrieval augmentation generation, and external data source integration, hpcGPT reduces the complexity and effort required to align information and identify dependencies between questions, answers, and the supporting information. This is particularly beneficial for research groups and computing centers with diverse application requirements and limited staff. hpcGPT extends and translates a suite of Cyberinfrastructure building blocks and technologies such as large language model training and inference service hosting.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.