Hi,

I was trying some stuff out on Cython 3.0, and I saw a bunch of errors of the form:

...
  warning: sklearn/metrics/_pairwise_distances_reduction/_radius_neighbors.pyx:954:49: Implicit noexcept declaration is deprecated. Function declaration should contain 'noexcept' keyword.
  warning: sklearn/metrics/_pairwise_distances_reduction/_radius_neighbors.pyx:966:11: Implicit noexcept declaration is deprecated. Function declaration should contain 'noexcept' keyword.
  warning: sklearn/utils/_sorting.pxd:9:7: Implicit noexcept declaration is deprecated. Function declaration should contain 'noexcept' keyword.
  warning: sklearn/utils/_vector_sentinel.pxd:12:60: Implicit noexcept declaration is deprecated. Function declaration should contain 'noexcept' keyword.

It seems that anytime nogil is used, the noexcept keyword must be typed. Is this of interest to patch?

Proposed Solution

It is thus time to replace the following instances in sklearn to be compatible with Cython3.0+:

• nogil with noexcept nogil and
• nogil except -1 with except -1 nogil

By running:

ag "nogil:" -c

You can see the following files are affected. The proposed strategy is to then tackle 1 submodule at a time (or one set of files at a time).

• sklearn/tree/_utils.pyx:16
• sklearn/tree/_oblique_tree.pyx:2
• sklearn/tree/_splitter.pyx:31
• sklearn/tree/_criterion.pyx:35
• sklearn/tree/_oblique_splitter.pyx:10
• sklearn/tree/_tree.pyx:19
• sklearn/metrics/_dist_metrics.pyx.tp:3
• sklearn/metrics/_pairwise_distances_reduction/_middle_term_computer.pyx.tp:17
• sklearn/metrics/_pairwise_distances_reduction/_base.pyx.tp:14
• sklearn/metrics/_pairwise_distances_reduction/_datasets_pair.pyx.tp:20
• sklearn/metrics/_pairwise_distances_reduction/_radius_neighbors.pyx.tp:16
• sklearn/metrics/_pairwise_distances_reduction/_argkmin.pyx.tp:16
• sklearn/metrics/_pairwise_fast.pyx:1
• sklearn/ensemble/_hist_gradient_boosting/histogram.pyx:8
• sklearn/ensemble/_hist_gradient_boosting/_predictor.pyx:2
• sklearn/ensemble/_hist_gradient_boosting/_bitset.pyx:5
• sklearn/ensemble/_hist_gradient_boosting/splitting.pyx:11
• sklearn/ensemble/_gradient_boosting.pyx:1
• sklearn/cluster/_k_means_elkan.pyx:2
• sklearn/cluster/_k_means_common.pyx:2
• sklearn/cluster/_k_means_minibatch.pyx:2
• sklearn/cluster/_k_means_lloyd.pyx:2
• sklearn/_loss/_loss.pyx.tp:37
• sklearn/_isotonic.pyx:1
• sklearn/linear_model/_sag_fast.pyx.tp:10
• sklearn/linear_model/_cd_fast.pyx:10
• sklearn/linear_model/_sgd_fast.pyx:26
• sklearn/utils/sparsefuncs_fast.pyx:1
• sklearn/utils/_random.pxd:1
• sklearn/utils/_openmp_helpers.pyx:1
• sklearn/utils/_sorting.pyx:2
• sklearn/utils/_cython_blas.pyx:11
• sklearn/utils/_weight_vector.pyx.tp:6
• sklearn/utils/_isfinite.pyx:4
• sklearn/utils/_seq_dataset.pyx.tp:8
• sklearn/utils/_heap.pyx:1
• sklearn/svm/_liblinear.pyx:1
• sklearn/svm/_libsvm.pyx:5
• sklearn/svm/_libsvm_sparse.pyx:3
• sklearn/manifold/_barnes_hut_tsne.pyx:3
• sklearn/preprocessing/_csr_polynomial_expansion.pyx:3
• sklearn/decomposition/_cdnmf_fast.pyx:1
• sklearn/decomposition/_online_lda_fast.pyx:1
• sklearn/neighbors/_quad_tree.pyx:6
• sklearn/neighbors/_binary_tree.pxi:3

You can run specifically ag ") nogil:" -Q, if you want a more fine-grained search.

Once the changes are made, one must run the following:

pip install --upgrade cython
cython --version
> # should be 3.0
pip install --verbose --editable . 2>&1 | grep -C 3 "sklearn.metrics"

and copy/paste the output for each PR to verify that they have indeed made the changes correctly.