During June 2025, contributed to the pytorch/torchtune repository by implementing a flexible activation mechanism within the GroupedExperts class. This work involved refactoring the activation logic to replace the previously hardcoded SiLU function with a pluggable self.act_fn, allowing for easier experimentation with different activation functions. The approach maintained backward compatibility while enhancing the modularity and adaptability of the codebase. Utilizing Python and leveraging deep learning frameworks such as PyTorch, the changes streamlined model tuning workflows and reduced future maintenance overhead. This feature improved the extensibility of activation logic, supporting more efficient development and experimentation in machine learning projects.