Worked on the intel/torch-xpu-ops repository to enhance XPU attention compatibility by refactoring the scaled dot product attention logic for improved PyTorch integration. The approach involved moving the sdp_choice logic to the PyTorch layer and introducing a device support stub, which reduced cross-backend branching and simplified maintenance. By removing the unimplemented sdpa_mem fallback and streamlining backend paths, the work improved both efficiency and maintainability of the attention mechanism in the XPU context. Utilized C++ and YAML alongside GPU programming and machine learning expertise to deliver a cleaner, more future-proof codebase, with a focus on code quality and integration alignment.