Contributed performance optimizations to the jeejeelee/vllm repository, focusing on MXFP8 GEMM kernels for AMD ROCm gfx950 hardware. Developed shape-adaptive tile selection and grid-swizzling techniques to enhance cache locality and occupancy for both dense-linear and grouped-MoE computation paths. Leveraged Python, PyTorch, and Triton to implement these GPU optimizations, targeting improved efficiency in matrix operations. Introduced a comprehensive test suite that validates the grouped GEMM implementation against a PyTorch reference, ensuring correctness and reliability. All work was delivered as a traceable commit under the ROCm performance initiative, reflecting a methodical approach to high-performance kernel engineering and validation.