Worked on the jeejeelee/vllm repository to deliver configurable performance enhancements for ROCm-based machine learning workloads. Developed environment flags that allow users to disable dynamic MXFP4 quantization and enable AITER-tuned GEMMs specifically for attention projection layers, providing greater flexibility in performance tuning and deployment portability. The implementation focused on GPU programming and quantization techniques using Python, targeting improved configurability and observability for attention-heavy models. By enabling dynamic control over quantization and GEMM strategies, the work addressed the need for adaptable performance optimization in diverse ROCm environments, contributing to more efficient and customizable software development for machine learning applications.