Developed a performance optimization feature for the modular/modular repository by introducing the RMSNormFused layer tailored for the DeepSeekV3.2 architecture. This work focused on fusing normalization with optional residual addition, enabling faster inference in production deep learning pipelines. The implementation provided distinct execution paths for scenarios with and without residuals, utilizing float32 precision and output casting as needed. Comprehensive unit tests were created to validate the interface across all use cases, ensuring reproducibility through explicit commit references. The project leveraged Python and deep learning techniques, contributing to the optimization strategy for machine learning workloads within modular’s architectural framework.