Worked on the ultralytics/ultralytics repository to address a critical robustness issue in the AAttn attention module, focusing on deep learning and neural networks using PyTorch and Python. The main contribution involved fixing a bug that caused crashes when the input dimension was not divisible by the number of attention heads, which previously limited the module’s flexibility. By carefully adjusting tensor reshaping and dimension calculations, the update ensured stable processing across a wide range of input sizes and batch configurations. This improvement enhanced the reliability of inference workflows, supporting more diverse datasets and deployment scenarios without compromising model stability.