This PR is tackling to the same issue with a quite old PR #13275 .

Edit: As @thomasjfox said, let's wait for @ogrisel review.

thomasjfox is especially unrelated to the scikit-learn project ;)

-> wrong guy

I did a quick benchmark and noticed that this PR leads to a runtime regression:

from sklearn.decomposition import LatentDirichletAllocation
from sklearn.datasets import make_multilabel_classification
from time import perf_counter

X, _ = make_multilabel_classification(random_state=0, n_samples=2_000,
                                      n_features=20)
lda = LatentDirichletAllocation(n_components=5, random_state=0)

start = perf_counter()
lda.fit(X)
delta = perf_counter() - start
print(f"{delta} s")

With this PR I get 7.3 s and on main I get 4.4 s. We would need to investigate before merging.

Thanks! I have reproduce the same slow down issue. I will investigate it.

You might want to try using a profiler such as viztracer, scalene or py-spy to pinpoint the part of the code responsible for the slowdown.

I identified and fixed the regression if PR #24528

Thanks @takoika, your contribution has been included in #24528