Developed a Triton-based row-wise FP8 to BF16 dequantization kernel for the ROCm/FBGEMM repository, focusing on enhancing the efficiency and flexibility of FP8 operations within the FBGEMM library. The work involved implementing a custom GPU kernel using Python and C++, with comprehensive unit tests to validate correctness across multiple tensor shapes and dimensions. By integrating this feature, the developer enabled broader adoption of FP8 formats and supported future optimizations in performance-critical workflows. The contribution also improved code quality and test coverage, aligning with ongoing efforts to integrate Triton and optimize GPU programming for high-performance machine learning applications.