MUMBAI, India, Jan. 2 -- Intellectual Property India has published a patent application (202541123851 A) filed by Malla Reddy (MR) Deemed to be University; Malla Reddy Vishwavidyapeeth; Malla Reddy University; Malla Reddy Engineering College For Women; and Malla Reddy College Of Engineering And Technology, Hyderabad, Telangana, on Dec. 9, 2025, for 'cross platform ai model trainer with adaptive optimization algorithms.'
Inventor(s) include Mr. P. Uday; Mrs. Kotte Shivani; Dr. Parthasaradhi Mayasala; Mr. A. Damodar; and Dr P Hari Krishna.
The application for the patent was published on Jan. 2, under issue no. 01/2026.
According to the abstract released by the Intellectual Property India: "The current invention reveals a universal Cross-Platform Artificial Intelligence (AI) Model Trainer that would help to smoothly train the deep learning models in the heterogeneous hardware settings. The modern machine learning process is often hampered by hardware-related dependencies, with models that are trained on high-performance server-grade Graphics Processing Units (GPUs) not working on, or working poorly with, edge devices, Central Processing Units (CPUs) or other accelerator architectures. These limitations are addressed through the invention of a hardware-agnostic abstraction layer, which translates high-level tensor operations into machine code specific to the underlying compute resources, whether CUDA-enabled hardware or OpenCL compatible hardware or proprietary Neural Processing Units (NPUs). In addition, the system incorporates a dynamic optimization engine which autonomously characterizes the computing resources (capabilities and constraints) of the host hardware in real-time. In contrast to old-fashioned training regimes that employ fixed hyperparameters (i.e. fixed batch size or learning rate), the suggested invention keeps track of the metrics of the system, such as memory bandwidth, thermal headroom, and floating-point operations per second (FLOPS). According to this telemetry, the system automatically scales training parameters dynamically in real time to avoid memory overflows and thermal throttling to ensure a maximum throughput with minimum manual intervention on the part of the data scientist. The invention also presents a distributed synchronization protocol that can coordinate the training of models with a group of different devices that have different processing speeds. With traditional distributed training, the slowest node (so called straggler problem) can be a bottleneck and slow down the cluster. The current mechanism alleviates this with an asynchronous gradient update system mixed with load balancing through weights, in which chunks of data are allocated to each device according to the current throughput of the device. This enables a work station, a laptop and a cloud server to work together in training a single model effectively. To conclude, the invention revealed gives a single framework that separates the design of the models and the execution of them in hardware. The system will tremendously decrease the development lifecycle of AI applications by automating the complexities of hardware interfacing and hyperparameter tuning, making model training more accessible and democratic by allowing commodity hardware to be used, and optimizing the use of resources in fragmented computing environments."
Disclaimer: Curated by HT Syndication.
