Developed a high-performance Cute-DSL backend for grouped-query attention decoding on Blackwell GPUs within the flashinfer-ai/flashinfer repository. Leveraged CUDA, GPU programming, and Python to enable efficient decoding with support for multiple data types, speculative execution, and advanced memory layouts. Modernized API surfaces and integrated new wrappers to streamline user access and improve runtime safety. Expanded test and benchmark coverage using PyTorch and pytest to validate correctness and throughput across diverse configurations. Addressed forward compatibility by migrating to updated APIs and fixed a critical buffer corruption issue, enhancing production reliability and enabling measurable throughput improvements in machine learning workloads.