Challenges and Standardization in PyTorch Ecosystem...

Lightning Talk: Challenges and Standardization in PyTorch Ecosystem Accelerators - Zesheng Zong, Huawei & Ashok Emani, Intel This talk aims to provide a practical and structured overview of the current challenges and recent standardization efforts in supporting accelerators within the PyTorch ecosystem. By sharing community-driven solutions and practical implementations, we hope to help more hardware developers and contributors effectively integrate with PyTorch and improve the overall developer experience. The TAC Out-of-Tree Working Group was formed to address these ecosystem gaps. Efforts include: Testing Framework & Quality Standards A public testing repository(pytorch-fdn/oota) was established to support daily CI testing for plugin stability. This initiative helps ensure that hardware plugins remain compatible with PyTorch core changes. Additionally, we standardized testing and quality standards for out-of-tree accelerators. Simplified Plugin and Feature Integration We optimized the integration mechanism and introduced automatic plugin loading to reduce the complexity of adding new devices. Including improving the code generalization and less coupling with CUDA, which previously made device-specific changes difficult. And make torch.compile better support integrating out-of-tree devices.