Worked on enhancing the InputObserver component in the pytorch/pytorch repository to improve robustness for models with dynamic input tensors. Addressed challenges faced by models like Gemma3, which require support for varying input signatures across inference steps, by implementing custom empty tensor handling and extending tolerance for missing inputs. Leveraged Python and PyTorch to introduce targeted tests that validate dynamic input scenarios, ensuring reliability and preventing regressions. Focused on model optimization and ONNX export compatibility, the work improved deployment reliability and maintainability for adaptive models. Emphasized thorough testing and documentation to support ongoing development and integration within the codebase.