scikit-learn · ogrisel · Feb 24, 2022 · Feb 23, 2022 · Feb 23, 2022 · Feb 23, 2022
diff --git a/doc/glossary.rst b/doc/glossary.rst
@@ -1555,6 +1555,9 @@ functions or non-estimator constructors.
         generally be interpreted as ``n_jobs=1``, unless the current
         :class:`joblib.Parallel` backend context specifies otherwise.
 
+        Note that even if ``n_jobs=1``, low-level parallelism (via Numpy and OpenMP)
+        might be used in some configuration.
+
         For more details on the use of ``joblib`` and its interactions with
         scikit-learn, please refer to our :ref:`parallelism notes
         <parallelism>`.

diff --git a/sklearn/metrics/_pairwise_distances_reduction.pyx b/sklearn/metrics/_pairwise_distances_reduction.pyx
@@ -112,14 +112,6 @@ cdef class PairwiseDistancesReduction:
         scikit-learn configuration for `pairwise_dist_chunk_size`,
         and use 256 if it is not set.
 
-    n_threads: int, default=None
-        The number of OpenMP threads to use for the reduction.
-        Parallelism is done on chunks and the sharding of chunks
-        depends on the `strategy` set on :method:`~PairwiseDistancesReduction.compute`.
-
-        See _openmp_effective_n_threads, for details about
-        the specification of n_threads.
-
     strategy : str, {'auto', 'parallel_on_X', 'parallel_on_Y'}, default=None
         The chunking strategy defining which dataset parallelization are made on.
 
@@ -220,7 +212,6 @@ cdef class PairwiseDistancesReduction:
         self,
         DatasetsPair datasets_pair,
         chunk_size=None,
-        n_threads=None,
         strategy=None,
      ):
         cdef:
@@ -231,7 +222,7 @@ cdef class PairwiseDistancesReduction:
 
         self.chunk_size = check_scalar(chunk_size, "chunk_size", Integral, min_val=20)
 
-        self.effective_n_threads = _openmp_effective_n_threads(n_threads)
+        self.effective_n_threads = _openmp_effective_n_threads()
 
         self.datasets_pair = datasets_pair
 
@@ -512,15 +503,6 @@ cdef class PairwiseDistancesArgKmin(PairwiseDistancesReduction):
         scikit-learn configuration for `pairwise_dist_chunk_size`,
         and use 256 if it is not set.
 
-    n_threads: int, default=None
-        The number of OpenMP threads to use for the reduction.
-        Parallelism is done on chunks and the sharding of chunks
-        depends on the `strategy` set on
-        :meth:`~PairwiseDistancesArgKmin.compute`.
-
-        See _openmp_effective_n_threads, for details about
-        the specification of n_threads.
-
     k: int, default=1
         The k for the argkmin reduction.
     """
@@ -544,7 +526,6 @@ cdef class PairwiseDistancesArgKmin(PairwiseDistancesReduction):
         str metric="euclidean",
         chunk_size=None,
         dict metric_kwargs=None,
-        n_threads=None,
         str strategy=None,
         bint return_distance=False,
     ):
@@ -574,15 +555,6 @@ cdef class PairwiseDistancesArgKmin(PairwiseDistancesReduction):
         metric_kwargs : dict, default=None
             Keyword arguments to pass to specified metric function.
 
-        n_threads : int, default=None
-            The number of OpenMP threads to use for the reduction.
-            Parallelism is done on chunks and the sharding of chunks
-            depends on the `strategy` set on
-            :meth:`~PairwiseDistancesArgKmin.compute`.
-
-            See _openmp_effective_n_threads, for details about
-            the specification of n_threads.
-
         strategy : str, {'auto', 'parallel_on_X', 'parallel_on_Y'}, default=None
             The chunking strategy defining which dataset parallelization are made on.
 
@@ -686,14 +658,12 @@ cdef class PairwiseDistancesArgKmin(PairwiseDistancesReduction):
         self,
         DatasetsPair datasets_pair,
         chunk_size=None,
-        n_threads=None,
         strategy=None,
         ITYPE_t k=1,
     ):
         super().__init__(
             datasets_pair=datasets_pair,
             chunk_size=chunk_size,
-            n_threads=n_threads,
             strategy=strategy,
         )
         self.k = check_scalar(k, "k", Integral, min_val=1)
@@ -932,7 +902,6 @@ cdef class FastEuclideanPairwiseDistancesArgKmin(PairwiseDistancesArgKmin):
         ITYPE_t k,
         bint use_squared_distances=False,
         chunk_size=None,
-        n_threads=None,
         strategy=None,
         metric_kwargs=None,
     ):
@@ -948,7 +917,6 @@ cdef class FastEuclideanPairwiseDistancesArgKmin(PairwiseDistancesArgKmin):
             # The datasets pair here is used for exact distances computations
             datasets_pair=DatasetsPair.get_for(X, Y, metric="euclidean"),
             chunk_size=chunk_size,
-            n_threads=n_threads,
             strategy=strategy,
             k=k,
         )

diff --git a/sklearn/metrics/tests/test_pairwise_distances_reduction.py b/sklearn/metrics/tests/test_pairwise_distances_reduction.py
@@ -1,5 +1,6 @@
 import numpy as np
 import pytest
+import threadpoolctl
 from numpy.testing import assert_array_equal, assert_allclose
 from scipy.sparse import csr_matrix
 from scipy.spatial.distance import cdist
@@ -229,9 +230,10 @@ def test_n_threads_agnosticism(
         return_distance=True,
     )
 
-    dist, indices = PairwiseDistancesReduction.compute(
-        X, Y, parameter, n_threads=1, return_distance=True
-    )
+    with threadpoolctl.threadpool_limits(limits=1, user_api="openmp"):
+        dist, indices = PairwiseDistancesReduction.compute(
+            X, Y, parameter, return_distance=True
+        )
 
     ASSERT_RESULT[PairwiseDistancesReduction](ref_dist, dist, ref_indices, indices)
 

diff --git a/sklearn/neighbors/_base.py b/sklearn/neighbors/_base.py
@@ -764,7 +764,6 @@ class from an array representing our data set and ask who's
                 k=n_neighbors,
                 metric=self.effective_metric_,
                 metric_kwargs=self.effective_metric_params_,
-                n_threads=self.n_jobs,
                 strategy="auto",
                 return_distance=return_distance,
             )