Worked on enhancing the NVIDIA/Megatron-LM repository by developing a memory-optimization feature for Mamba model inference. Introduced fine-grained activation offloading, allowing selective offloading of activation tensors to improve memory efficiency during large-scale inference. Implemented a centralized preprocessing method to manage offloading parameters and integrated safeguards to prevent offloading when the feature is disabled, ensuring stable operation across configurations. Validated the solution by measuring memory footprint and stability, which enabled support for larger batch sizes with predictable latency in production environments. The work leveraged deep learning and model optimization techniques, utilizing Python to address scalability and memory management challenges.