Tensor Shapes (pyrefly.org)

Pyrefly has introduced an experimental feature that tracks tensor shapes in PyTorch models, significantly enhancing the development experience by providing end-to-end static type checking for shape transformations. This feature automates the inlay type hints that display the shape of intermediate tensors, allowing developers to easily identify shape mismatches and avoid errors that could lead to incorrect results without relying on print statements or executing the code. With Pyrefly, users can now see shapes like Tensor[B, T, NEmbedding] for embeddings in real-time, improving overall debugging efficiency. The underlying technology combines symbolic integer arithmetic with a library of shape transform specifications tailored to PyTorch operators, enabling inference of tensor shapes even from minimal annotations. This approach stands out by prioritizing ease of use and less verbose syntax as compared to previous attempts, making it more accessible for developers working with complex neural networks. Additionally, Pyrefly allows extending shape coverage for new PyTorch operators through a dedicated DSL, facilitating active community contributions. This innovation is set to streamline the workflow of AI/ML practitioners, fostering more robust model development and reducing the risk of shape-related bugs.