Worked on the pytorch/pytorch repository to deliver ROCshMEM Triton backend integration, enabling symmetric memory support on ROCm and unblocking previously skipped NVSHMEM Triton tests. Developed a unified backend dispatch mechanism and consolidated shared utilities using Python, focusing on distributed systems and GPU programming. Refactored SHMEM tests into a backend-agnostic framework, introducing architecture-aware bindings and device bitcode discovery to broaden hardware compatibility. The approach improved maintainability, test coverage, and developer experience by centralizing entry points and reducing code duplication. All changes were validated through comprehensive test suites, demonstrating robust cross-backend support and enhancing the overall reliability of the codebase.