Worked on the pytorch/pytorch repository to enhance the backward graph copy pipeline, focusing on enabling robust cross-device support for autograd in fused modules. Leveraged deep learning and machine learning expertise with PyTorch and Python to implement call_module support in copy_paste_aot_backward_graph, ensuring consistent behavior across CPU and HPU environments. Addressed error handling for tensor indexing within backward-graph copy paths, reducing failure modes during model export and import. Improved gradient retention for non-leaf tensors, preserving correct gradient flow in complex, multi-module models. The work increased portability and reliability for production models operating in mixed hardware environments without introducing new bugs.