ci/test: patch unstable test_bleu_score_functional (#2533)

(cherry picked from commit d76b82e)

tests/unittests/_helpers/wrappers.py

@@ -6,6 +6,12 @@
 
 ALLOW_SKIP_IF_OUT_OF_MEMORY = os.getenv("ALLOW_SKIP_IF_OUT_OF_MEMORY", "0") == "1"
 ALLOW_SKIP_IF_BAD_CONNECTION = os.getenv("ALLOW_SKIP_IF_BAD_CONNECTION", "0") == "1"
+_ERROR_CONNECTION_MESSAGE_PATTERNS = (
+    "We couldn't connect to",
+    "Connection error",
+    "Can't load",
+    "`nltk` resource `punkt` is",
+)
 
 
 def skip_on_running_out_of_memory(reason: str = "Skipping test as it ran out of memory."):
@@ -33,18 +39,25 @@ def skip_on_connection_issues(reason: str = "Unable to load checkpoints from Hug
     The tests run normally if no connection issue arises, and they're marked as skipped otherwise.
 
     """
-    _error_msg_starts = ["We couldn't connect to", "Connection error", "Can't load", "`nltk` resource `punkt` is"]
 
     def test_decorator(function: Callable, *args: Any, **kwargs: Any) -> Optional[Callable]:
         @wraps(function)
         def run_test(*args: Any, **kwargs: Any) -> Optional[Any]:
+            from urllib.error import URLError
+
             try:
                 return function(*args, **kwargs)
+            except URLError as ex:
+                if "Error 403: Forbidden" not in str(ex) or not ALLOW_SKIP_IF_BAD_CONNECTION:
+                    raise ex
+                pytest.skip(reason)
             except (OSError, ValueError) as ex:
-                if all(msg_start not in str(ex) for msg_start in _error_msg_starts):
+                if (
+                    all(msg_start not in str(ex) for msg_start in _ERROR_CONNECTION_MESSAGE_PATTERNS)
+                    or not ALLOW_SKIP_IF_BAD_CONNECTION
+                ):
                     raise ex
-                if ALLOW_SKIP_IF_BAD_CONNECTION:
-                    pytest.skip(reason)
+                pytest.skip(reason)
 
         return run_test
 

tests/unittests/text/test_sacre_bleu.py

@@ -21,6 +21,7 @@
 from torchmetrics.functional.text.sacre_bleu import AVAILABLE_TOKENIZERS, _TokenizersLiteral, sacre_bleu_score
 from torchmetrics.text.sacre_bleu import SacreBLEUScore
 
+from unittests._helpers import skip_on_connection_issues
 from unittests.text._helpers import TextTester
 from unittests.text._inputs import _inputs_multiple_references
 
@@ -69,6 +70,7 @@ def test_bleu_score_class(self, ddp, preds, targets, tokenize, lowercase):
             metric_args=metric_args,
         )
 
+    @skip_on_connection_issues(reason="could not download model or tokenizer")
     def test_bleu_score_functional(self, preds, targets, tokenize, lowercase):
         """Test functional implementation of metric."""
         if _should_skip_tokenizer(tokenize):