PyTorch has announced the adoption of Pyrefly to enhance type checking within its core repository and in several ecosystem projects, including Helion, TorchTitan, and Ignite. This transition aims to ensure greater consistency and prevent common bugs that often go unnoticed in dynamic code.

Why Pyrefly?

The choice of Pyrefly is motivated by the need for a type checker that is fast, easy to use, consistent across different development environments, and actively maintained. Pyrefly offers a balance between speed and accuracy, simplifies configuration, and ensures consistency between IDEs, CLIs, and continuous integration (CI) environments.

Performance

Benchmarks have highlighted a significant performance improvement: type checking PyTorch with Pyrefly (v44.1) takes only 5.5 seconds, compared to the 50.6 seconds required with MyPy. This speed increase allows contributors to identify bugs more quickly and identify gaps in type coverage.

Simplified Configuration

The migration to Pyrefly has made it possible to switch to a single unified configuration, simplifying the management of type coverage and strictness levels in the code. Previously, using MyPy required managing multiple configuration files.

Consistency across environments

Pyrefly ensures consistent results between IDEs, CLIs, and CI, eliminating the discrepancies that occurred with the use of different type checking engines. This translates into a more predictable and less frustrating development experience.

Additional benefits for contributors

The adoption of Pyrefly has led to an improvement in code quality, thanks to the identification and resolution of numerous bugs. Native integration with various IDEs offers real-time feedback, documentation, and diagnostics directly in the editor. Advanced typing features allow you to develop safer and more expressive code.

Future projects

PyTorch plans to continue collaborating with the Pyrefly team to further improve typing in the code. The goal is to strengthen type annotations to enable the community to use PyTorch in production environments more securely.