Another fix

Author: mattdangerw

keras_nlp/src/models/pali_gemma/pali_gemma_backbone_test.py

@@ -85,7 +85,7 @@ def test_saved_model(self):
             input_data=self.input_data,
         )
 
-    @pytest.mark.large
+    @pytest.mark.extra_large
     def test_smallest_preset(self):
         self.run_preset_test(
             cls=PaliGemmaBackbone,

keras_nlp/src/utils/tensor_utils.py

@@ -137,21 +137,23 @@ def convert_preprocessing_inputs(x):
         if numpy_x.dtype == np.int64:
             numpy_x = numpy_x.astype(np.int32)
         # We have non-ragged, non-string input. Use backbend type.
-        return ops.convert_to_tensor(numpy_x)
+        x = ops.convert_to_tensor(numpy_x)
+        # Torch will complain about device placement for GPU tensors.
+        if keras.config.backend() == "torch":
+            x = x.cpu()
+        return x
     if is_tensor_type(x):
         # String or ragged types we keep as tf.
         if isinstance(x, tf.RaggedTensor) or x.dtype == tf.string:
             return x
         # If we have a string input, use tf.tensor.
         if isinstance(x, np.ndarray) and x.dtype.type is np.str_:
             return tf.convert_to_tensor(x)
+        x = ops.convert_to_tensor(x)
         # Torch will complain about device placement for GPU tensors.
         if keras.config.backend() == "torch":
-            import torch
-
-            if isinstance(x, torch.Tensor):
-                x = x.cpu()
-        return ops.convert_to_tensor(x)
+            x = x.cpu()
+        return x
     return x
 
 

keras_nlp/src/utils/transformers/preset_loader.py

@@ -71,3 +71,7 @@ def load_backbone(self, cls, load_weights, **kwargs):
 
     def load_tokenizer(self, cls, **kwargs):
         return self.converter.convert_tokenizer(cls, self.preset, **kwargs)
+
+    def load_image_converter(self, cls, **kwargs):
+        # TODO: auto resize.
+        return None