diff --git a/src/scib_metrics/_nmi_ari.py b/src/scib_metrics/_nmi_ari.py
index 715fe0b..47ada34 100644
--- a/src/scib_metrics/_nmi_ari.py
+++ b/src/scib_metrics/_nmi_ari.py
@@ -8,8 +8,7 @@
 from sklearn.metrics.cluster import adjusted_rand_score, normalized_mutual_info_score
 from sklearn.utils import check_array
 
-from scib_metrics import settings
-
+from ._settings import warnings_stacklevel
 from .utils import KMeans, check_square
 
 logger = logging.getLogger(__name__)
@@ -115,7 +114,7 @@ def nmi_ari_cluster_labels_leiden(
             warnings.warn(
                 "Using for loop over clustering resolutions. `pip install joblib` for parallelization.",
                 UserWarning,
-                stacklevel=settings.warnings_stacklevel,
+                stacklevel=warnings_stacklevel,
             )
             out = [_compute_nmi_ari_cluster_labels(X, labels, r) for r in resolutions]
         nmi_ari = np.array(out)
diff --git a/src/scib_metrics/_pcr_comparison.py b/src/scib_metrics/_pcr_comparison.py
index 2637519..f8e1b83 100644
--- a/src/scib_metrics/_pcr_comparison.py
+++ b/src/scib_metrics/_pcr_comparison.py
@@ -1,7 +1,6 @@
 import warnings
 
-from scib_metrics import settings
-
+from ._settings import warnings_stacklevel
 from ._types import NdArray
 from .utils import principal_component_regression
 
@@ -52,7 +51,7 @@ def pcr_comparison(
                 "PCR comparison score is negative, meaning variance contribution "
                 "increased after integration. Setting to 0.",
                 UserWarning,
-                stacklevel=settings.warnings_stacklevel,
+                stacklevel=warnings_stacklevel,
             )
             pcr_compared = 0
     else:
diff --git a/src/scib_metrics/benchmark/_core.py b/src/scib_metrics/benchmark/_core.py
index f9bc12a..8069069 100644
--- a/src/scib_metrics/benchmark/_core.py
+++ b/src/scib_metrics/benchmark/_core.py
@@ -18,7 +18,7 @@
 from tqdm import tqdm
 
 import scib_metrics
-from scib_metrics import settings
+from scib_metrics._settings import warnings_stacklevel
 from scib_metrics.nearest_neighbors import NeighborsOutput, pynndescent
 
 Kwargs = Dict[str, Any]
@@ -205,7 +205,7 @@ def benchmark(self) -> None:
             warnings.warn(
                 "The benchmark has already been run. Running it again will overwrite the previous results.",
                 UserWarning,
-                stacklevel=settings.warnings_stacklevel,
+                stacklevel=warnings_stacklevel,
             )
 
         if not self._prepared: