Developed GPU memory benchmarking enhancements for the Intel-tensorflow/tensorflow and openxla/xla repositories, focusing on enabling peak memory stat resets for PJRT devices. Leveraging C++ and GPU programming expertise, introduced a workflow that allows developers to benchmark GPU memory usage without running the full profiler, reducing overhead and accelerating performance tuning cycles. The implementation included targeted GPU tests on NVIDIA 3090s and incorporated code hygiene improvements to ensure stability and maintainability. By improving memory-usage visibility and supporting efficient benchmarking, this work enables teams to iterate more quickly on memory-related optimizations, with careful attention to testing and validation throughout the process.