Checks on a Pull Request

When you open a pull request on 🤗 Transformers, a fair number of checks will be run to make sure the patch you are adding is not breaking anything existing. Those checks are of four types:

• regular tests
• documentation build
• code and documentation style
• general repository consistency

In this document, we will take a stab at explaining what those various checks are and the reason behind them, as well as how to debug them locally if one of them fails on your PR.

Note that, ideally, they require you to have a dev install:

pip install transformers[dev]

or for an editable install:

pip install -e .[dev]

inside the Transformers repo. Since the number of optional dependencies of Transformers has grown a lot, it’s possible you don’t manage to get all of them. If the dev install fails, make sure to install the Deep Learning framework you are working with (PyTorch, TensorFlow and/or Flax) then do

pip install transformers[quality]

or for an editable install:

pip install -e .[quality]

Tests

All the jobs that begin with ci/circleci: run_tests_ run parts of the Transformers testing suite. Each of those jobs focuses on a part of the library in a certain environment: for instance ci/circleci: run_tests_pipelines_tf runs the pipelines test in an environment where TensorFlow only is installed.

Note that to avoid running tests when there is no real change in the modules they are testing, only part of the test suite is run each time: a utility is run to determine the differences in the library between before and after the PR (what GitHub shows you in the “Files changes” tab) and picks the tests impacted by that diff. That utility can be run locally with:

python utils/tests_fetcher.py

from the root of the Transformers repo. It will:

1. Check for each file in the diff if the changes are in the code or only in comments or docstrings. Only the files with real code changes are kept.
2. Build an internal map that gives for each file of the source code of the library all the files it recursively impacts. Module A is said to impact module B if module B imports module A. For the recursive impact, we need a chain of modules going from module A to module B in which each module imports the previous one.
3. Apply this map on the files gathered in step 1, which gives us the list of model files impacted by the PR.
4. Map each of those files to their corresponding test file(s) and get the list of tests to run.

When executing the script locally, you should get the results of step 1, 3 and 4 printed and thus know which tests are run. The script will also create a file named test_list.txt which contains the list of tests to run, and you can run them locally with the following command:

python -m pytest -n 8 --dist=loadfile -rA -s $(cat test_list.txt)

Just in case anything slipped through the cracks, the full test suite is also run daily.

Documentation build

The build_pr_documentation job builds and generates a preview of the documentation to make sure everything looks okay once your PR is merged. A bot will add a link to preview the documentation in your PR. Any changes you make to the PR are automatically updated in the preview. If the documentation fails to build, click on Details next to the failed job to see where things went wrong. Often, the error is as simple as a missing file in the toctree.

If you’re interested in building or previewing the documentation locally, take a look at the README.md in the docs folder.

Code and documentation style

Code formatting is applied to all the source files, the examples and the tests using black and ruff. We also have a custom tool taking care of the formatting of docstrings and rst files (utils/style_doc.py), as well as the order of the lazy imports performed in the Transformers __init__.py files (utils/custom_init_isort.py). All of this can be launched by executing

make style

The CI checks those have been applied inside the ci/circleci: check_code_quality check. It also runs ruff, that will have a basic look at your code and will complain if it finds an undefined variable, or one that is not used. To run that check locally, use

make quality

This can take a lot of time, so to run the same thing on only the files you modified in the current branch, run

make fixup

This last command will also run all the additional checks for the repository consistency. Let’s have a look at them.

Repository consistency

This regroups all the tests to make sure your PR leaves the repository in a good state, and is performed by the ci/circleci: check_repository_consistency check. You can locally run that check by executing the following:

make repo-consistency

This checks that:

• All objects added to the init are documented (performed by utils/check_repo.py)
• All __init__.py files have the same content in their two sections (performed by utils/check_inits.py)
• All code identified as a copy from another module is consistent with the original (performed by utils/check_copies.py)
• All configuration classes have at least one valid checkpoint mentioned in their docstrings (performed by utils/check_config_docstrings.py)
• All configuration classes only contain attributes that are used in corresponding modeling files (performed by utils/check_config_attributes.py)
• The translations of the READMEs and the index of the doc have the same model list as the main README (performed by utils/check_copies.py)
• The auto-generated tables in the documentation are up to date (performed by utils/check_table.py)
• The library has all objects available even if not all optional dependencies are installed (performed by utils/check_dummies.py)

Should this check fail, the first two items require manual fixing, the last four can be fixed automatically for you by running the command

make fix-copies

Additional checks concern PRs that add new models, mainly that:

• All models added are in an Auto-mapping (performed by utils/check_repo.py)
• All models are properly tested (performed by utils/check_repo.py)