Developed an adaptive backend selector for FlashAttention-3 within the PriorLabs/TabPFN repository, targeting the v3 architecture on Hopper GPUs. This feature automatically determines and applies the most efficient attention backend based on available hardware and input parameters, optimizing performance for large-scale deep learning models. The implementation leveraged GPU programming and PyTorch, focusing on performance optimization and seamless integration with existing workflows. No major bugs were addressed during this period, as stability of the v3 integration was maintained. The work involved collaborative development and demonstrated a strong understanding of deep learning infrastructure and hardware-aware software design in Python.