Worked on the pytorch/pytorch repository to address a memory-format regression in the index_select decomposition that impacted torch.roll operations. Using Python and leveraging expertise in memory management and tensor operations, implemented a targeted fix to ensure memory-format preservation when creating contiguous tensors. This change improved correctness and predictability for channel-last workflows and reduced unexpected performance variations across memory formats. Enhanced the testing suite by adding a regression test to reproduce and verify the issue, and expanded coverage for memory-format edge cases. The work focused on maintaining consistent behavior and performance in PyTorch’s tensor manipulation functions through careful testing and code updates.