Worked on the ggml-org/llama.cpp repository to deliver a targeted performance optimization for Nemotron Nano v2, focusing on enabling CUDA Graph usage to streamline memory copy operations and reduce overall runtime. Leveraged C++ and CUDA to integrate graph-based execution, which improved throughput and lowered latency for CUDA workloads on the target hardware. The approach maintained compatibility with Nemotron Nano v2, ensuring seamless deployment for edge inference scenarios. This work demonstrated skills in GPU programming, performance engineering, and cross-hardware optimization, while laying the foundation for future enhancements in graph-based GPU performance tuning within the llama.cpp codebase. No bugs were addressed.