Developed and integrated end-to-end performance testing for the Stable Diffusion 3.5 Medium model within the vllm-omni repository, focusing on validating configuration-driven performance improvements. Leveraged Python and CI/CD pipelines to automate test execution and ensure reliable benchmarking as part of the continuous integration workflow. Expanded test coverage to include performance-critical inference paths, providing actionable visibility into model behavior under various configurations. This approach enabled faster optimization cycles and supported more reliable release processes. The work emphasized robust testing practices and machine learning model evaluation, establishing a foundation for ongoing performance monitoring and guiding future enhancements to the vllm-omni project.