MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202621082006 A) filed by Symbiosis International Deemed University on July 02, 2026, for Predictive Gpu Cluster Resource Provisioning System Using Foundation Model Based Demand Forecasting And Fragmentation-Aware Scheduling.
Inventors include Karan Misal; and Dr. Bhupesh Kumar Dewangan.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: ABSTRACT PREDICTIVE GPU CLUSTER RESOURCE PROVISIONING SYSTEM USING FOUNDATION MODEL BASED DEMAND FORECASTING AND FRAGMENTATION-AWARE SCHEDULING The present invention discloses a predictive GPU cluster resource provisioning system (100) comprising a data ingestion and preprocessing pipeline (110), a historical demand time series database (120), a time series foundation model inference engine (130) employing a pre-trained Mixture-of-Experts transformer architecture, a forecast-to-distribution converter (140), a dynamic TypicalPods generator (150), an asynchronous forecast service (160), and a fragmentation-aware scheduling engine (170). The system (100) ingests pod lifecycle events from a container orchestration platform, generates aggregate GPU demand time series, and applies zero-shot multi-step forecasting using a foundation model pre-trained on over 300 billion time series tokens. Forecasts are converted into dynamic pod arrival distributions that replace static historical profiles in a modified Fragmentation Gradient Descent scheduler. The system achieves a zero-shot mean absolute error of 0.0123 on GPU demand forecasting and reduces GPU fragmentation by 22 to 30 percent compared to conventional static scheduling approaches. [
Disclaimer: Curated by HT Syndication.
