Thomas J. Fan

For PyPI, we do not use secrets anymore and use [Trusted Publisher](https://docs.pypi.org/trusted-publishers/) As for uploads to `scientific-python-nightly-wheels` or the staging index, we should create an GitHub Actions environment for them.

I opened https://github.com/scikit-learn/scikit-learn/pull/28441 to add an environment for the `upload_anaconda` secrets.

> where does the third-campus-393023:scikit-learn-gpu-v4 machine image come from? This is a custom image on my GCP account with conda + nvidia drivers installed. This makes start up times much...

> It seems that you did not finish your sentence. I reordered the paragraphs and did not remove the last part in the editing. 😅

I did not use a `Dockerfile` to create the custom image on GCP. I had a running instance, installed the Nvidia drivers + conda, and then created a clone. I...

This was an real failure with 32bit Linux, when `SKLEARN_TESTS_GLOBAL_RANDOM_SEED="50"`.

Just in case the logs are removed, here is the trackback: Traceback ```python-traceback def test_inverse_transform( whiten, n_components, expected_mixing_shape, global_random_seed, global_dtype ): # Test FastICA.inverse_transform n_samples = 100 rng = np.random.RandomState(global_random_seed)...

Looks like another error with global random state: ```python loss = 'absolute_error', subsample = 0.5, global_random_seed = 9 @pytest.mark.parametrize("loss", ("squared_error", "absolute_error", "huber")) @pytest.mark.parametrize("subsample", (1.0, 0.5)) def test_regression_dataset(loss, subsample, global_random_seed): #...

We recently updated the `black` version on `main`. Sync up this PR with `main`, so that this PR picks up the new black version in the CI.

As a default, I think two approvals is too much, because there are some simple PRs such as documentation or CI fixes that should only require one review. I'll say...