A bitwise reproducible deep learning framework (github.com)

Researchers have introduced RepDL, a deep learning framework that ensures bitwise reproducibility across different hardware platforms during training and inference. This specialized library addresses a significant challenge in the AI/ML community: the inconsistencies in results when running identical models on various devices, which can arise from non-deterministic operations within frameworks like PyTorch. By implementing reproducible operations, RepDL allows developers to achieve consistent results, enhancing reliability in experimental settings. RepDL is designed for academic and non-production use, providing users with a set of reproducible functions and modules compatible with PyTorch. It includes a straightforward installation process and enables reproducible inference with minimal code adjustments. Notably, the library highlights specific operations that, while traditionally non-reproducible in PyTorch, can yield consistent results when using customized RepDL implementations. The project invites community contributions, encouraging developers to extend its capabilities and further enhance reproducibility in deep learning research. This initiative could significantly impact the validation of experimental results in AI and machine learning, fostering a more reliable research environment.