GitLab

This should stabilize formatting.

This will improve version coverage.

Take advantage of this to simplify the Circle CI configuration.

Don't bother with tensorboardX: it's a fallback for PyTorch < 1.1.0.

requirements.txt isn't up to date.

This prevents transformers from being importable simply because the CWD
is the root of the git repository, while not being importable from other
directories. That led to inconsistent behavior, especially in examples.

Once you fetch this commit, in your dev environment, you must run:

    $ pip uninstall transformers
    $ pip install -e .

These libraries aren't always installed in the virtual environment where
isort is running. Declaring them properly avoids mixing these
third-party imports with local imports.

We need https://github.com/timothycrosley/isort/pull/1000 but there's no
release with this fix yet, so we'll install from GitHub.

This optimizes the total run time of the Circle CI test suite.

We're already using as many processes in parallel as we have CPU cores.
Furthermore, the number of core may be incorrectly calculated as 36
(we've seen this in pytest-xdist) which make compound the problem.

PyTorch performance craters without this.

This should prevent two issues:

- hitting API rate limits for tests that hit the HF API
- multiplying the cost of expensive test setups

Set the number of CPUs manually based on the Circle CI resource class,
or else we're getting 36 CPUs, which is far too much (perhaps that's
the underlying hardware and not what Circle CI allocates to us).

Don't parallelize the custom tokenizers tests because they take less
than one second to run and parallelization actually makes them slower.

We add a script and a CI workflow to check that all download links
present in the source code are valid.