Riccardo Felluga addressed a memory-format regression in the PyTorch repository, focusing on the index_select decomposition that impacted torch.roll operations. Using Python and leveraging expertise in memory management and tensor operations, Riccardo implemented a fix that ensures memory-format preservation when creating contiguous tensors, thereby improving correctness and predictability for channel-last workflows. The solution included updating torch._refs.index_select to suggest the appropriate memory format and adding a targeted regression test to verify the fix. By enhancing test coverage for memory-format edge cases, Riccardo’s work provided a deeper level of reliability and consistency for developers working with PyTorch tensor operations.