During August 2025, Znifly contributed to the pytorch/pytorch repository by developing a dynamic backend parameter for the MTIA framework’s GPU vertical pass and pattern matcher. This feature, implemented in Python with a focus on GPU programming and backend development, allows runtime selection of the backend while maintaining backward compatibility through the default 'inductor' backend. The approach enables greater flexibility for experimentation and performance tuning within the MTIA GPU path, supporting future roadmap goals. Znifly’s work established a robust foundation for backend extensibility, addressing the need for adaptable infrastructure without introducing breaking changes or regressions to existing workflows.