During February 2026, Katoue contributed to the pytorch/pytorch repository by optimizing symbolic shape computations in PyTorch. Katoue focused on the SymInt copy constructor and assignment operator, redesigning them in C++ to directly copy data and manually increment reference counts. This approach eliminated unnecessary temporary allocations and redundant reference count cycles, reducing memory overhead and improving performance for models using dynamic shapes. The work demonstrated careful attention to low-level memory management and performance optimization, addressing a critical path in the framework. Katoue’s changes were technically rigorous, reflecting a deep understanding of C++ and the complexities of efficient resource handling in large codebases.