Developed a performance-focused feature for the flashinfer-ai/flashinfer repository, introducing a heuristic-based configuration selector for the CuTe-DSL FP4 GEMM kernel. This solution, implemented in Python with a focus on algorithm design and GPU programming, enables device-aware selection of tile and cluster shapes to optimize throughput when autotuning is disabled. By caching configuration decisions and adjusting for small-K scenarios, the approach reduced setup overhead and improved consistency across diverse hardware, including B200 and GB300 systems. The work was validated on a wide range of GEMM workloads, with code quality maintained through updated tests and successful pre-commit checks.