Developed an optional rotary position embedding backend for microsoft/DeepSpeed, integrating torchembed to enable fused operations via Triton when available and performance criteria are met. The implementation included a robust ImportError guard to ensure compatibility in environments lacking torchembed, and expanded packaging to expose the backend through deepspeed[torchembed]. Comprehensive unit tests were added to validate numerical correctness and gradient flow across various sequence lengths and dimensions, following established patterns from transformers and vLLM. The work leveraged Python, CUDA, and PyTorch, focusing on backend optimization and deep learning infrastructure, with thorough validation and clear commit documentation for maintainability.