Fix BertScore on GPU

CHANGELOG.md

@@ -130,6 +130,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed compatibility of `ClasswiseWrapper` with the `prefix` argument of `MetricCollection` ([#843](https://github.com/PyTorchLightning/metrics/pull/843))
 
 
+- Fixed `BestScore` on GPU ([#912](https://github.com/PyTorchLightning/metrics/pull/912))
+
+
 ## [0.7.3] - 2022-03-23
 
 ### Fixed

tests/text/test_bertscore.py

@@ -94,16 +94,20 @@ def test_score_fn_with_idf(preds, targets):
     "preds,targets",
     [(preds, targets)],
 )
+@pytest.mark.parametrize("device", ["cpu", "cuda"])
 @pytest.mark.skipif(not _BERTSCORE_AVAILABLE, reason="test requires bert_score")
-def test_score_fn_all_layers(preds, targets):
+def test_score_fn_all_layers(preds, targets, device):
     """Tests for functional and all layers."""
+    if not torch.cuda.is_available() and device == "cuda":
+        pytest.skip("Test requires GPU support")
+
     original_score = original_bert_score(
         preds, targets, model_type=MODEL_NAME, all_layers=True, idf=False, batch_size=3
     )
     original_score = _parse_original_bert_score(original_score)
 
     metrics_score = metrics_bert_score(
-        preds, targets, model_name_or_path=MODEL_NAME, all_layers=True, idf=False, batch_size=3
+        preds, targets, model_name_or_path=MODEL_NAME, all_layers=True, idf=False, batch_size=3, device=device
     )
 
     for metric in _METRICS:

torchmetrics/functional/text/bert.py

@@ -337,7 +337,7 @@ def _get_embeddings_and_idf_scale(
             batch["input_ids_idf"] * processed_attention_mask if idf else processed_attention_mask.type(out.dtype)
         )
         input_ids_idf /= input_ids_idf.sum(-1, keepdim=True)
-        idf_scale_list.append(input_ids_idf)
+        idf_scale_list.append(input_ids_idf.cpu())
 
     embeddings = torch.cat(embeddings_list)
     idf_scale = torch.cat(idf_scale_list)