Worked on enhancing evaluation robustness and performance in PyTorch-based repositories, focusing on deep learning and neural network workflows. In pytorch/torchtune, addressed evaluation failures in mixed-precision models by introducing an exclusion list for non-checkable parameter types, which improved reliability across diverse configurations and pruned vocabularies. Also resolved loading issues for vocab-pruned models during evaluation, reducing pipeline disruptions. In pytorch/executorch, implemented YOCO (You Only Cache Once) key/value sharing in the attention layer, allowing reuse of projections and improving transformer throughput. Leveraged Python, PyTorch, and machine learning expertise to deliver targeted solutions that stabilized evaluation and optimized attention computation.