Developed the Flextron Elasticity feature for the NVIDIA/Megatron-LM repository, enabling dynamic routing and masking to support on-demand resource adaptation in large-scale deep learning models. The work involved designing and integrating elasticity management hooks and configuration options across model components, allowing for improved throughput and resource utilization while maintaining model quality. Implemented primarily in Python, the solution leveraged expertise in PyTorch, distributed systems, and model optimization. Integration was carefully aligned with existing Megatron-LM pipelines, and the changes were prepared for upstream review. The contribution focused on feature delivery, with minor integration adjustments and no major bug fixes within the scope.