Apoorv Gupta enhanced the apple/axlearn repository by developing rematerialization patterns aimed at improving memory efficiency and training scalability for transformer-based models. Using Python and leveraging deep learning and machine learning expertise, Apoorv implemented remat strategies for neuron configurations, enabling selective offloading of transformer layers to reduce peak memory usage during training. The technical approach included updating regex patterns for saving and offloading model components, as well as expanding test coverage to ensure robust integration of remat within the training loop. This work provided a solid foundation for reducing memory footprint and increasing throughput in large-scale model training scenarios.