Developed and integrated CUDA Profiler API support into the kvcache-ai/sglang repository, enabling precise GPU benchmarking for single-batch workloads. This work introduced profiling start and stop hooks, along with new command-line options, allowing targeted and granular performance analysis. Leveraging Python scripting and CUDA programming, the implementation provided detailed metrics on execution time and GPU utilization, laying the groundwork for data-driven optimization of GPU workloads. The focus remained on feature delivery rather than bug fixes, with comprehensive documentation to support future enhancements. This instrumentation established a robust foundation for ongoing performance profiling and optimization within the sglang project’s GPU pipeline.