facebookresearch/hydra

Hydra is a Python framework from Meta's research team that makes it easier to manage configuration in complex applications. Configuration refers to all the settings and options your program needs to run: things like which model to use, where to find data, what learning rate to apply, or which server to connect to. As applications grow, managing all these settings in a clean and flexible way becomes a real challenge. Hydra is built to solve that. The core idea is that you define your configuration in structured files (using a format called YAML), and Hydra loads and composes those files when your program starts. You can override any setting from the command line without changing the files, which makes it easy to run experiments with different parameters. You can also combine multiple configuration files together, so a large application can split its settings across many files that get merged at runtime. Hydra is especially popular in machine learning research, where people frequently run the same program many times with slightly different settings to compare results. It reduces the boilerplate involved in parsing command-line arguments and managing configuration across runs. A number of third-party libraries build on top of it, including templates that combine Hydra with popular training frameworks for convenience. Installation is a single pip command. The stable version is 1.3, and a development version (1.4) is available from source. It supports Python 3.10 through 3.14 and is licensed under the MIT open-source license. Documentation lives on the project's own website, and the community uses GitHub Discussions and StackOverflow for questions.