Worked on the modelscope/data-juicer repository to address stability and resource efficiency in image processing workflows. Focused on resolving a memory leak in the ImageNSFWFilter by applying torch.no_grad() during model inference, which reduced unnecessary memory usage and prevented gradient tracking in production inference scenarios. This change improved the deterministic behavior of memory consumption, allowing for longer-running batch jobs without unexpected spikes. The solution was implemented in Python and leveraged deep learning frameworks, with particular attention to memory management and performance optimization. All modifications were validated to ensure no impact on end-user features, and the fix was prepared for production deployment.