Developed a reliability-focused feature for the pytorch/pytorch repository by implementing deterministic model evaluation algorithms on ROCm-enabled devices. This work involved removing outdated and flaky models from the accuracy check process, which reduced test noise and stabilized continuous integration outcomes. Leveraging Python and expertise in data analysis and machine learning, the developer established a deterministic evaluation path that improved both the reliability and performance of model comparisons. The approach aligned the ROCm backend with deterministic algorithms, enabling faster and more trustworthy end-to-end model evaluation. The contribution addressed reliability challenges in model assessment workflows without introducing new bugs during the development period.