[BugFix] Proper auto-batch size for unbatched tensors

tensordict/_td.py

@@ -62,6 +62,7 @@
     _NON_STR_KEY_ERR,
     _NON_STR_KEY_TUPLE_ERR,
     _parse_to,
+    _pass_through,
     _prune_selected_keys,
     _set_item,
     _set_max_batch_size,
@@ -3809,7 +3810,7 @@ def _get_non_tensor(self, key: NestedKey, default=NO_DEFAULT):
     def _get_str(self, key, default):
         if key in self.keys() and _is_tensor_collection(self.entry_class(key)):
             data = self._source._get_str(key, NO_DEFAULT)
-            if is_non_tensor(data):
+            if _pass_through(data):
                 return data[self.idx]
             return _SubTensorDict(data, self.idx)
         return self._source._get_at_str(key, self.idx, default=default)

tensordict/utils.py

@@ -1788,7 +1788,7 @@ def _set_max_batch_size(source: T, batch_dims=None):
     """Updates a tensordict with its maximum batch size."""
     from tensordict.base import _is_tensor_collection
 
-    tensor_data = [val for val in source.values() if not is_non_tensor(val)]
+    tensor_data = [val for val in source.values() if not _pass_through(val)]
 
     for val in tensor_data:
         if _is_tensor_collection(type(val)):

test/test_tensordict.py

@@ -11813,6 +11813,13 @@ class SubTC(NonTensorData): ...
 
 
 class TestUnbatchedTensor:
+    def test_auto_batch_size(self):
+        td = TensorDict(a=UnbatchedTensor(0), b=torch.randn(2, 3)).auto_batch_size_(
+            batch_dims=2
+        )
+        assert td.shape == (2, 3)
+        assert td["a"] == 0
+
     def test_unbatched(self):
         assert UnbatchedTensor._pass_through
         td = TensorDict(