Worked on the llvm/torch-mlir repository to address a critical issue in the Fx Importer related to non-persistent buffer handling during experimental mutation workflows. The solution involved modifying the importer to retrieve buffer values from the module’s constants rather than the state_dict, which previously caused incorrect behavior when mutation support was enabled. This change improved the stability and correctness of mutation-enabled scenarios, ensuring that the workflow remains reliable. The work was implemented using Python and leveraged expertise in machine learning and PyTorch, demonstrating a focused approach to maintaining integration quality and supporting advanced features in the software development process.