Worked on the inclusionAI/AReaL repository to deliver a training loop performance optimization focused on deep learning workflows. Addressed inefficiencies in GPU and CPU memory residency by refining onload and offload sequences for model parameters, which reduced unnecessary data transfers and improved training throughput. The approach maintained model parameters in GPU memory during critical stages such as compute_values, ppo_update, and checkpointing, minimizing latency and residency churn. Validated these changes at scale using Python, PyTorch, and FSDP2 across multi-GPU setups, confirming both stability and scalability. The work demonstrated strong skills in performance optimization and large-scale machine learning system design.