Developed a precision-aware optimizer with decoupled gradients for the NVIDIA/Megatron-LM repository, focusing on enhancing distributed deep learning workflows. The solution introduced a configuration-driven approach to enable precision-aware optimization within Megatron-FSDP, allowing users to opt in via a single flag in the distributed training configuration. By leveraging PyTorch and distributed computing techniques, the work improved memory efficiency and scalability, supporting larger models and batch sizes without sacrificing convergence. Integration with existing mixed-precision workflows ensured compatibility with both FP16 and FP32 modes, while validation within Megatron-FSDP workflows confirmed robust performance and maintainability across diverse distributed training environments.