Worked on optimizing GPU tensor handling within the pytorch/TensorRT repository by delivering a memory-efficient improvement to the prepare_inputs workflow. Focused on reducing unnecessary duplication of GPU tensor data, the solution enhanced memory efficiency during model parameter handling, which supports more stable inference and the potential for larger batch sizes. The approach involved careful manipulation of tensors and integration with existing GPU programming practices, all implemented in Python. Emphasized performance optimization and code quality, with thorough unit testing to ensure reliability. This work addressed a key bottleneck in scalable inference, contributing to more efficient resource utilization in deep learning pipelines.