Worked on core kernel acceleration improvements for the DarkLight1337/vllm repository, focusing on enhancing GPU matrix operations. Upgraded the CUTLASS library to version 3.8, integrating the latest performance and stability features into the build system using CMake and C++. Developed initializers for multiple CUTLASS epilogue variants, enabling configurable post-processing for dense matrix computations and laying the groundwork for future GPU kernel optimizations. Linked these enhancements to the continuous integration pipeline, ensuring reliable deployment and testing. This work improved the potential performance and flexibility of inference paths in large language model workloads, supporting ongoing development of high-performance GPU programming solutions.