Ensure tensors are at least 1d for pad and concat (huggingface#17179)

* Ensure tensors are at least 1d for pad and concat

* Compatibility

* Fix

* Fix

* Add test

* Retrigger CI

* Consistency with master

* Retrigger CI

src/transformers/trainer_pt_utils.py

@@ -55,8 +55,22 @@
 logger = logging.get_logger(__name__)
 
 
+def atleast_1d(tensor_or_array: Union[torch.Tensor, np.ndarray]):
+    if isinstance(tensor_or_array, torch.Tensor):
+        if hasattr(torch, "atleast_1d"):
+            tensor_or_array = torch.atleast_1d(tensor_or_array)
+        elif tensor_or_array.ndim < 1:
+            tensor_or_array = tensor_or_array[None]
+    else:
+        tensor_or_array = np.atleast_1d(tensor_or_array)
+    return tensor_or_array
+
+
 def torch_pad_and_concatenate(tensor1, tensor2, padding_index=-100):
     """Concatenates `tensor1` and `tensor2` on first axis, applying padding on the second if necessary."""
+    tensor1 = atleast_1d(tensor1)
+    tensor2 = atleast_1d(tensor2)
+
     if len(tensor1.shape) == 1 or tensor1.shape[1] == tensor2.shape[1]:
         return torch.cat((tensor1, tensor2), dim=0)
 
@@ -72,6 +86,9 @@ def torch_pad_and_concatenate(tensor1, tensor2, padding_index=-100):
 
 def numpy_pad_and_concatenate(array1, array2, padding_index=-100):
     """Concatenates `array1` and `array2` on first axis, applying padding on the second if necessary."""
+    array1 = atleast_1d(array1)
+    array2 = atleast_1d(array2)
+
     if len(array1.shape) == 1 or array1.shape[1] == array2.shape[1]:
         return np.concatenate((array1, array2), axis=0)
 
@@ -149,8 +166,7 @@ def nested_xla_mesh_reduce(tensors, name):
 
         if isinstance(tensors, (list, tuple)):
             return type(tensors)(nested_xla_mesh_reduce(t, f"{name}_{i}") for i, t in enumerate(tensors))
-        if tensors.ndim == 0:
-            tensors = tensors[None]
+        tensors = atleast_1d(tensors)
         return xm.mesh_reduce(name, tensors, torch.cat)
     else:
         raise ImportError("Torch xla must be installed to use `nested_xla_mesh_reduce`")
@@ -160,8 +176,7 @@ def distributed_concat(tensor: Any, num_total_examples: Optional[int] = None) ->
     try:
         if isinstance(tensor, (tuple, list)):
             return type(tensor)(distributed_concat(t, num_total_examples) for t in tensor)
-        if len(tensor.shape) <= 0:
-            tensor = tensor[None]
+        tensor = atleast_1d(tensor)
         output_tensors = [tensor.clone() for _ in range(dist.get_world_size())]
         dist.all_gather(output_tensors, tensor)
         concat = torch.cat(output_tensors, dim=0)
@@ -1031,7 +1046,7 @@ def smp_gather(tensor):
                 f"Can't gather the values of type {type(tensor)}, only of nested list/tuple/dicts of tensors."
             )
         all_tensors = smp.allgather(tensor, smp.CommGroup.DP_GROUP)
-        all_tensors = [t if len(t.shape) > 0 else t[None] for t in all_tensors]
+        all_tensors = [atleast_1d(t) for t in all_tensors]
         return torch.cat([t.cpu() for t in all_tensors], dim=0)
 
     def smp_nested_concat(tensor):

tests/trainer/test_trainer_utils.py

@@ -41,6 +41,8 @@
         SequentialDistributedSampler,
         ShardSampler,
         get_parameter_names,
+        numpy_pad_and_concatenate,
+        torch_pad_and_concatenate,
     )
 
     class TstLayer(nn.Module):
@@ -459,6 +461,18 @@ def mock_training_loop_function(batch_size):
             mock_training_loop_function()
             self.assertEqual("CUDA out of memory", cm.args[0])
 
+    def test_pad_and_concatenate_with_1d(self):
+        """Tests whether pad_and_concatenate works with scalars."""
+        array1 = 1.0
+        array2 = 2.0
+        result = numpy_pad_and_concatenate(array1, array2)
+        self.assertTrue(np.array_equal(np.array([1.0, 2.0]), result))
+
+        tensor1 = torch.tensor(1.0)
+        tensor2 = torch.tensor(2.0)
+        result = torch_pad_and_concatenate(tensor1, tensor2)
+        self.assertTrue(torch.equal(result, torch.Tensor([1.0, 2.0])))
+
     def test_remove_columns_collator(self):
         class MockLogger:
             def __init__(self) -> None: