Ivan Fedorov contributed to the pytorch/executorch and pytorch/torchtune repositories, focusing on deep learning model efficiency and evaluation robustness. He developed YOCO KV sharing in the executorch attention layer, enabling layers to reuse key and value projections from donor layers, which reduced redundant computation and improved transformer throughput. In torchtune, Ivan enhanced mixed-precision model evaluation by introducing an exclusion list for non-checkable parameter types, stabilizing evaluation across diverse configurations and pruned vocabularies. His work, primarily in Python and PyTorch, demonstrated a strong understanding of neural network internals and addressed both performance and reliability in complex machine learning pipelines.