Developed CUDNN normalization support for GPT-OSS-20B training within the NVIDIA-NeMo/Megatron-Bridge repository, focusing on enhancing pretraining throughput and resource utilization for large-scale deep learning models. The work involved integrating cuDNN normalization into the training path, leveraging Python and CUDA to optimize performance during model pretraining. By enabling TE cudnn norm with gpt_oss_20b, the implementation targeted improved efficiency in machine learning workflows without introducing major bug fixes during the period. This contribution demonstrated a strong emphasis on performance optimization and effective use of deep learning frameworks, addressing the computational demands of modern language model training environments.