Worked on enhancing GPU compatibility for the Marlin MoE kernel in the jeejeelee/vllm repository by introducing support for SM 12.x CUDA architectures. Leveraged CMake and CUDA to update the default architecture lists in alignment with the CUDA compiler version, ensuring that the build process accurately reflects available hardware capabilities. Implemented runtime GPU capability checks to prevent operations from running on unsupported devices, thereby reducing potential runtime errors and broadening the range of compatible hardware. This feature-focused contribution improved the reliability and flexibility of Marlin MoE workloads, demonstrating depth in GPU programming and cross-team collaboration within a complex codebase.