Worked on optimizing the Titan Training Framework in the huggingface/torchtitan repository by migrating its training workflow to a DCP ZOC-based approach for LLAMA3 8B. This involved replacing the default asynchronous and pinned memory model to achieve more efficient training and stable resource utilization. Enhanced checkpoint management and strengthened asynchronous operations to streamline workflows and reduce wait times. All modifications were consolidated into a single, auditable commit for clear traceability. The work leveraged deep learning and distributed systems expertise, utilizing Python and PyTorch to address performance bottlenecks and improve reliability in large-scale machine learning model training environments.