Worked on the ROCm/aiter repository to address GPU memory management challenges in the MoE CI pipeline, focusing on preventing out-of-memory errors during test runs. Applied explicit GPU memory management techniques using CUDA and PyTorch, introducing garbage collection and cache clearing after each test case to minimize memory fragmentation. This approach stabilized CI tests in resource-constrained environments, reducing flaky runs and improving overall reliability. By resolving a critical OOM issue, the work enhanced the CI feedback loop, enabling faster iteration for GPU-based models. Demonstrated strong skills in Python and testing, with a focus on practical solutions for large-scale machine learning workflows.