Worked on enhancing the reliability and correctness of RotaryEmbedding within the JustinTong0323/sglang repository, focusing on deep learning models deployed across diverse GPU hardware. Addressed a critical bug in the FP32 computation path by refining the logic for applying cached cosine and sine values, which previously led to numerical inaccuracies on different devices. This targeted fix improved cross-platform stability and ensured consistent model accuracy in production environments. Utilized Python and PyTorch to implement and document the solution, emphasizing maintainability and robust deployment. The work demonstrated a strong understanding of GPU computing and numerical precision in deep learning infrastructure.