Worked on the jeejeelee/vllm repository to address stability and correctness in CUDA workflows, focusing on compatibility with Blackwell desktop GPUs. The primary contribution involved fixing a bug in CMake configuration that preserved architecture-specific CUDA suffixes for SM12x, which previously led to NaN issues in CUDA operations. By ensuring the correct handling of CUDA architecture versioning, the update improved reliability and reduced computational errors on targeted hardware. The work demonstrated attention to detail in bug fixing and collaboration, utilizing CMake and CUDA expertise to enhance the robustness of the build process without introducing new user-facing features during the development period.