Developed a targeted training optimization feature for the pytorch/torchtune repository, focusing on improving single-device full finetuning workflows. Designed and implemented a learning rate scheduler using Python and PyTorch, enabling more efficient training convergence and reducing overall wall-clock time for experiments on constrained hardware. Integrated the scheduler directly into the finetuning pipeline, enhancing both stability and efficiency during model experimentation. This work leveraged deep learning and model optimization expertise to address the challenges of resource-limited training environments. The contribution aligned with torchtune’s efficiency-focused roadmap, providing a practical solution for accelerating experimentation without compromising model performance or training reliability.