Fix device scope issues (#1841)

We want to always place tf ops on a GPU device, this broke.

keras_nlp/src/tokenizers/byte_pair_tokenizer.py

@@ -540,6 +540,7 @@ def tokenize(self, inputs):
         if self.add_prefix_space:
             inputs = tf.strings.join([" ", inputs])
 
+        inputs = tf.convert_to_tensor(inputs)
         unbatched = inputs.shape.rank == 0
         if unbatched:
             inputs = tf.expand_dims(inputs, 0)

keras_nlp/src/tokenizers/sentence_piece_tokenizer.py

@@ -238,6 +238,7 @@ def _check_vocabulary(self):
     @preprocessing_function
     def tokenize(self, inputs):
         self._check_vocabulary()
+        inputs = tf.convert_to_tensor(inputs)
         unbatched = inputs.shape.rank == 0
         if unbatched:
             inputs = tf.expand_dims(inputs, 0)

keras_nlp/src/tokenizers/word_piece_tokenizer.py

@@ -473,6 +473,7 @@ def _check_vocabulary(self):
     @preprocessing_function
     def tokenize(self, inputs):
         self._check_vocabulary()
+        inputs = tf.convert_to_tensor(inputs)
         unbatched = inputs.shape.rank == 0
         pattern = None
         if self.split and self.special_tokens_in_strings:

keras_nlp/src/utils/tensor_utils.py

@@ -53,28 +53,29 @@ def preprocessing_function(fn):
 
     params = inspect.signature(fn).parameters
     accepts_labels = all(k in params for k in ("x", "y", "sample_weight"))
-    with tf.device("cpu"):
-        if not accepts_labels:
+    if not accepts_labels:
 
-            @functools.wraps(fn)
-            def wrapper(self, x, **kwargs):
+        @functools.wraps(fn)
+        def wrapper(self, x, **kwargs):
+            with tf.device("cpu"):
                 x = convert_preprocessing_inputs(x)
                 with no_convert_scope():
                     x = fn(self, x, **kwargs)
                 return convert_preprocessing_outputs(x)
 
-        else:
+    else:
 
-            @functools.wraps(fn)
-            def wrapper(self, x, y=None, sample_weight=None, **kwargs):
+        @functools.wraps(fn)
+        def wrapper(self, x, y=None, sample_weight=None, **kwargs):
+            with tf.device("cpu"):
                 x, y, sample_weight = convert_preprocessing_inputs(
                     (x, y, sample_weight)
                 )
                 with no_convert_scope():
                     x = fn(self, x, y=y, sample_weight=sample_weight, **kwargs)
                 return convert_preprocessing_outputs(x)
 
-        return wrapper
+    return wrapper
 
 
 def convert_preprocessing_inputs(x):

keras_nlp/src/utils/tensor_utils_test.py

@@ -23,12 +23,13 @@
 from keras_nlp.src.utils.tensor_utils import convert_preprocessing_outputs
 from keras_nlp.src.utils.tensor_utils import convert_to_ragged_batch
 from keras_nlp.src.utils.tensor_utils import is_tensor_type
+from keras_nlp.src.utils.tensor_utils import preprocessing_function
 from keras_nlp.src.utils.tensor_utils import tensor_to_list
 
 
 class ConvertHelpers(TestCase):
     def test_basics(self):
-        inputs = ops.array([1, 2, 3])
+        inputs = [1, 2, 3]
         # Convert to tf.
         outputs = convert_preprocessing_inputs(inputs)
         self.assertAllEqual(outputs, ops.array(inputs))
@@ -92,6 +93,18 @@ def to_list(x):
         inputs = tree.flatten(tree.map_structure(to_list, inputs))
         self.assertAllEqual(outputs, inputs)
 
+    def test_placement(self):
+        # Make sure we always place preprocessing on the CPU on all backends.
+        @preprocessing_function
+        def test(self, inputs):
+            for x in inputs:
+                if isinstance(x, tf.Tensor):
+                    self.assertTrue("CPU" in x.device)
+                    self.assertFalse("GPU" in x.device)
+            return inputs
+
+        test(self, ([1, 2, 3], ["foo", "bar"], "foo"))
+
 
 class TensorToListTest(TestCase):
     def test_ragged_input(self):