Worked on the AI-Hypercomputer/torchprime repository to enhance the stability of TPU-based testing infrastructure. Addressed a persistent issue in the test_trainer by implementing dynamic batch sizing and mesh configuration that adapts to the actual TPU device count, ensuring that test harnesses accurately reflect available hardware. This approach reduced flaky test outcomes and improved the reliability of continuous integration runs. Leveraged Python, PyTorch, and XLA to align dummy datasets and mesh setups with hardware resources, strengthening reproducibility for distributed training scenarios. The work demonstrated a methodical approach to hardware-aware testing, focusing on robust, maintainable solutions for TPU environments without introducing new features.