diff --git a/CHANGELOG.md b/CHANGELOG.md
index a253f90f6ef..5d5c9554ef0 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -27,7 +27,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
--
+- Delete `Device2Host` caused by comm with device and host ([#2840](https://github.com/PyTorchLightning/metrics/pull/2840))
 
 
 ---
diff --git a/src/torchmetrics/functional/classification/precision_recall_curve.py b/src/torchmetrics/functional/classification/precision_recall_curve.py
index 3c5a840efa1..991606e4b25 100644
--- a/src/torchmetrics/functional/classification/precision_recall_curve.py
+++ b/src/torchmetrics/functional/classification/precision_recall_curve.py
@@ -182,8 +182,14 @@ def _binary_precision_recall_curve_format(
         preds = preds[idx]
         target = target[idx]
 
-    if not torch.all((preds >= 0) * (preds <= 1)):
-        preds = preds.sigmoid()
+    # "sigmoid_cpu" not implemented for 'Half'
+    if preds.dtype != torch.float16 or preds.device != torch.device("cpu"):
+        out_of_bounds = (preds < 0) | (preds > 1)
+        out_of_bounds = out_of_bounds.any()
+        preds = torch.where(out_of_bounds, preds.sigmoid(), preds)
+    else:
+        if not torch.all((preds >= 0) * (preds <= 1)):
+            preds = preds.sigmoid()
 
     thresholds = _adjust_threshold_arg(thresholds, preds.device)
     return preds, target, thresholds
@@ -761,8 +767,15 @@ def _multilabel_precision_recall_curve_format(
     """
     preds = preds.transpose(0, 1).reshape(num_labels, -1).T
     target = target.transpose(0, 1).reshape(num_labels, -1).T
-    if not torch.all((preds >= 0) * (preds <= 1)):
-        preds = preds.sigmoid()
+
+    # "sigmoid_cpu" not implemented for 'Half'
+    if preds.dtype != torch.float16 or preds.device != torch.device("cpu"):
+        out_of_bounds = (preds < 0) | (preds > 1)
+        out_of_bounds = out_of_bounds.any()
+        preds = torch.where(out_of_bounds, preds.sigmoid(), preds)
+    else:
+        if not torch.all((preds >= 0) * (preds <= 1)):
+            preds = preds.sigmoid()
 
     thresholds = _adjust_threshold_arg(thresholds, preds.device)
     if ignore_index is not None and thresholds is not None:
diff --git a/tests/unittests/classification/test_precision_recall_curve.py b/tests/unittests/classification/test_precision_recall_curve.py
index 7c034c528e6..3772073536c 100644
--- a/tests/unittests/classification/test_precision_recall_curve.py
+++ b/tests/unittests/classification/test_precision_recall_curve.py
@@ -106,15 +106,27 @@ def test_binary_precision_recall_curve_dtype_cpu(self, inputs, dtype):
         """Test dtype support of the metric on CPU."""
         preds, target = inputs
         if (preds < 0).any() and dtype == torch.half:
-            pytest.xfail(reason="torch.sigmoid in metric does not support cpu + half precision")
-        self.run_precision_test_cpu(
-            preds=preds,
-            target=target,
-            metric_module=BinaryPrecisionRecallCurve,
-            metric_functional=binary_precision_recall_curve,
-            metric_args={"thresholds": None},
-            dtype=dtype,
-        )
+            try:
+                self.run_precision_test_cpu(
+                    preds=preds,
+                    target=target,
+                    metric_module=BinaryPrecisionRecallCurve,
+                    metric_functional=binary_precision_recall_curve,
+                    metric_args={"thresholds": None},
+                    dtype=dtype,
+                )
+            except Exception as e:
+                print(f"An unexpected error occurred: {e}")
+                pytest.xfail(reason="torch.sigmoid in metric does not support cpu + half precision")
+        else:
+            self.run_precision_test_cpu(
+                preds=preds,
+                target=target,
+                metric_module=BinaryPrecisionRecallCurve,
+                metric_functional=binary_precision_recall_curve,
+                metric_args={"thresholds": None},
+                dtype=dtype,
+            )
 
     @pytest.mark.skipif(not torch.cuda.is_available(), reason="test requires cuda")
     @pytest.mark.parametrize("dtype", [torch.half, torch.double])