Worked on the NVIDIA-NeMo/Megatron-Bridge repository to deliver checkpoint swizzling for GLU weights in Megatron, targeting improved efficiency and flexibility in model checkpointing for large-scale deep learning training. Developed a fused gemm and swiglu kernel using Python and PyTorch, integrating directly with Megatron-LM to support the GLU path. This approach enhanced training resilience and throughput by enabling more robust and scalable checkpointing workflows. The work demonstrated expertise in deep learning, model checkpointing, and GPU kernel development, focusing on performance optimization and seamless integration with existing machine learning infrastructure. No major bugs were addressed during this period, emphasizing feature delivery.