init

transformers/tests/models/pixtral/test_image_processing_pixtral.py

@@ -0,0 +1,287 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+import pytest
+from packaging import version
+
+from transformers.image_utils import load_image
+from transformers.testing_utils import require_torch, require_torch_gpu, require_vision, slow, torch_device
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+from ...test_processing_common import url_to_local_path
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import PixtralImageProcessor
+
+    if is_torchvision_available():
+        from transformers import PixtralImageProcessorFast
+
+
+class PixtralImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        max_num_images_per_sample=3,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        patch_size=None,
+        do_normalize=True,
+        image_mean=[0.48145466, 0.4578275, 0.40821073],
+        image_std=[0.26862954, 0.26130258, 0.27577711],
+        do_convert_rgb=True,
+    ):
+        super().__init__()
+        size = size if size is not None else {"longest_edge": 24}
+        patch_size = patch_size if patch_size is not None else {"height": 8, "width": 8}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.max_num_images_per_sample = max_num_images_per_sample
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.patch_size = patch_size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "patch_size": self.patch_size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+        }
+
+    def expected_output_image_shape(self, images):
+        if not isinstance(images, (list, tuple)):
+            images = [images]
+
+        batch_size = len(images)
+        return_height, return_width = 0, 0
+        for image in images:
+            if isinstance(image, Image.Image):
+                width, height = image.size
+            elif isinstance(image, np.ndarray):
+                height, width = image.shape[:2]
+            elif isinstance(image, torch.Tensor):
+                height, width = image.shape[-2:]
+
+            max_height = max_width = self.size.get("longest_edge")
+
+            ratio = max(height / max_height, width / max_width)
+            if ratio > 1:
+                height = int(np.floor(height / ratio))
+                width = int(np.floor(width / ratio))
+
+            patch_height, patch_width = self.patch_size["height"], self.patch_size["width"]
+            num_height_tokens = (height - 1) // patch_height + 1
+            num_width_tokens = (width - 1) // patch_width + 1
+
+            return_height = max(num_height_tokens * patch_height, return_height)
+            return_width = max(num_width_tokens * patch_width, return_width)
+
+        return batch_size, self.num_channels, return_height, return_width
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        images = prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+        return images
+
+
+@require_torch
+@require_vision
+class PixtralImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = PixtralImageProcessor if is_vision_available() else None
+    fast_image_processing_class = PixtralImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = PixtralImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "patch_size"))
+            self.assertTrue(hasattr(image_processing, "do_rescale"))
+            self.assertTrue(hasattr(image_processing, "rescale_factor"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+            self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
+
+    # The following tests are overridden as PixtralImageProcessor can return images of different sizes
+    # and thus doesn't support returning batched tensors
+
+    def test_call_pil(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PIL images
+            image_inputs_list = self.image_processor_tester.prepare_image_inputs()
+            for image in image_inputs_list:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs_list[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list[0])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            encoded_images = image_processing(image_inputs_list, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list)
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs_list = self.image_processor_tester.prepare_image_inputs(numpify=True)
+            for image in image_inputs_list:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs_list[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list[0])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            batch_encoded_images = image_processing(image_inputs_list, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list)
+            self.assertEqual(tuple(batch_encoded_images.shape), expected_output_image_shape)
+
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs_list = self.image_processor_tester.prepare_image_inputs(torchify=True)
+            for image in image_inputs_list:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs_list[0], return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list[0])
+            self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
+
+            # Test batched
+            batch_encoded_images = image_processing(image_inputs_list, return_tensors="pt").pixel_values
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs_list)
+            self.assertEqual(tuple(batch_encoded_images.shape), expected_output_image_shape)
+
+    @require_vision
+    @require_torch
+    def test_slow_fast_equivalence(self):
+        dummy_image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
+
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        encoding_slow = image_processor_slow(dummy_image, return_tensors="pt")
+        encoding_fast = image_processor_fast(dummy_image, return_tensors="pt")
+        self._assert_slow_fast_tensors_equivalence(encoding_slow.pixel_values[0][0], encoding_fast.pixel_values[0][0])
+
+    @require_vision
+    @require_torch
+    def test_slow_fast_equivalence_batched(self):
+        dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
+            self.skipTest(
+                reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
+            )
+
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        encoding_slow = image_processor_slow(dummy_images, return_tensors="pt")
+        encoding_fast = image_processor_fast(dummy_images, return_tensors="pt")
+
+        for i in range(len(encoding_slow.pixel_values)):
+            self._assert_slow_fast_tensors_equivalence(
+                encoding_slow.pixel_values[i][0], encoding_fast.pixel_values[i][0]
+            )
+
+    @slow
+    @require_torch_gpu
+    @require_vision
+    @pytest.mark.torch_compile_test
+    def test_can_compile_fast_image_processor(self):
+        if self.fast_image_processing_class is None:
+            self.skipTest("Skipping compilation test as fast image processor is not defined")
+        if version.parse(torch.__version__) < version.parse("2.3"):
+            self.skipTest(reason="This test requires torch >= 2.3 to run.")
+
+        torch.compiler.reset()
+        input_image = torch.randint(0, 255, (3, 224, 224), dtype=torch.uint8)
+        image_processor = self.fast_image_processing_class(**self.image_processor_dict)
+        output_eager = image_processor(input_image, device=torch_device, return_tensors="pt")
+
+        image_processor = torch.compile(image_processor, mode="reduce-overhead")
+        output_compiled = image_processor(input_image, device=torch_device, return_tensors="pt")
+
+        self._assert_slow_fast_tensors_equivalence(
+            output_eager.pixel_values[0][0], output_compiled.pixel_values[0][0], atol=1e-4, rtol=1e-4, mean_atol=1e-5
+        )
+
+    @unittest.skip(reason="PixtralImageProcessor doesn't treat 4 channel PIL and numpy consistently yet")  # FIXME Amy
+    def test_call_numpy_4_channels(self):
+        pass

transformers/tests/models/pixtral/test_modeling_pixtral.py

@@ -0,0 +1,129 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch Pixtral model."""
+
+import unittest
+
+from transformers import (
+    PixtralVisionConfig,
+    PixtralVisionModel,
+    is_torch_available,
+)
+from transformers.testing_utils import (
+    require_torch,
+    torch_device,
+)
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor
+
+
+if is_torch_available():
+    import torch
+
+
+class PixtralVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        # in Pixtral, the seq length equals the number of patches * batch_size because the patches are flattened
+        self.seq_length = (image_size // patch_size) ** 2 * batch_size
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        image_sizes = torch.tensor(
+            [[self.image_size, self.image_size]] * self.batch_size, dtype=torch.long, device=torch_device
+        )
+        config = self.get_config()
+
+        return config, pixel_values, image_sizes
+
+    def get_config(self):
+        return PixtralVisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, image_sizes = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values, "image_sizes": image_sizes}
+        return config, inputs_dict
+
+
+@require_torch
+class PixtralVisionModelModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Model tester for `PixtralVisionModel`.
+    """
+
+    all_model_classes = (PixtralVisionModel,) if is_torch_available() else ()
+    additional_model_inputs = ["image_sizes"]
+    test_pruning = False
+    test_head_masking = False
+    test_torchscript = False
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = PixtralVisionModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=PixtralVisionConfig, has_text_modality=False)
+
+    def test_model_get_set_embeddings(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (torch.nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, torch.nn.Linear))

transformers/tests/models/pixtral/test_processing_pixtral.py

@@ -0,0 +1,271 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+
+from transformers.testing_utils import require_vision
+from transformers.utils import is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin, url_to_local_path
+
+
+if is_vision_available():
+    from transformers import PixtralProcessor
+
+
+@require_vision
+class PixtralProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = PixtralProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.url_0 = url_to_local_path(
+            "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg"
+        )
+        cls.image_0 = np.random.randint(255, size=(3, 876, 1300), dtype=np.uint8)
+        cls.url_1 = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        cls.image_1 = np.random.randint(255, size=(3, 480, 640), dtype=np.uint8)
+        cls.image_2 = np.random.randint(255, size=(3, 1024, 1024), dtype=np.uint8)
+
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+        processor = PixtralProcessor.from_pretrained("mistral-community/pixtral-12b")
+        processor.save_pretrained(self.tmpdirname)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def test_image_token_filling(self):
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        # Important to check with non square image
+        image = torch.randint(0, 2, (3, 500, 316))
+        expected_image_tokens = 640
+        image_token_index = 10
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image"},
+                    {"type": "text", "text": "What is shown in this image?"},
+                ],
+            },
+        ]
+        inputs = processor(
+            text=[processor.apply_chat_template(messages)],
+            images=[image],
+            return_tensors="pt",
+        )
+        image_tokens = (inputs["input_ids"] == image_token_index).sum().item()
+        self.assertEqual(expected_image_tokens, image_tokens)
+
+    def test_processor_with_single_image(self):
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        prompt_string = "USER: [IMG]\nWhat's the content of the image? ASSISTANT:"
+
+        # Make small for checking image token expansion
+        processor.image_processor.size = {"longest_edge": 30}
+        processor.image_processor.patch_size = {"height": 2, "width": 2}
+
+        # Test passing in an image
+        inputs_image = processor(text=prompt_string, images=self.image_0, return_tensors="pt")
+        self.assertIn("input_ids", inputs_image)
+        self.assertTrue(len(inputs_image["input_ids"]) == 1)
+        self.assertIsInstance(inputs_image["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([1, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_image["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to "USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the content of the image? ASSISTANT:"
+            [21510,  1058,  1032,    10,    10,    12,    10,    10,    13,  1010, 7493,  1681,  1278,  4701,  1307,  1278,  3937,  1063,  1349,  4290, 16002, 41150,  1058]
+        )
+        # fmt: on
+
+        # Test passing in a url
+        inputs_url = processor(text=prompt_string, images=self.url_0, return_tensors="pt")
+        self.assertIn("input_ids", inputs_url)
+        self.assertTrue(len(inputs_url["input_ids"]) == 1)
+        self.assertIsInstance(inputs_url["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([1, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_url["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to "USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the content of the image? ASSISTANT:"
+            [21510,  1058,  1032,    10,    10,    12,    10,    10,    13,  1010, 7493,  1681,  1278,  4701,  1307,  1278,  3937,  1063,  1349,  4290, 16002, 41150,  1058]
+        )
+        # fmt: on
+
+        # Test passing inputs as a single list
+        inputs_image = processor(text=prompt_string, images=[self.image_0], return_tensors="pt")
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([1, 3, 32, 32]))
+
+        # fmt: off
+        self.assertEqual(
+            inputs_image["input_ids"][0].tolist(),
+            [21510,  1058,  1032,    10,    10,    12,    10,    10,    13,  1010, 7493,  1681,  1278,  4701,  1307,  1278,  3937,  1063,  1349,  4290, 16002, 41150,  1058]
+        )
+        # fmt: on
+
+        # Test as nested single list
+        inputs_image = processor(text=prompt_string, images=[[self.image_0]], return_tensors="pt")
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([1, 3, 32, 32]))
+
+        # fmt: off
+        self.assertEqual(
+            inputs_image["input_ids"][0].tolist(),
+            [21510,  1058,  1032,    10,    10,    12,    10,    10,    13,  1010, 7493,  1681,  1278,  4701,  1307,  1278,  3937,  1063,  1349,  4290, 16002, 41150,  1058]
+        )
+        # fmt: on
+
+    def test_processor_with_multiple_images_single_list(self):
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        prompt_string = "USER: [IMG][IMG]\nWhat's the difference between these two images? ASSISTANT:"
+
+        # Make small for checking image token expansion
+        processor.image_processor.size = {"longest_edge": 30}
+        processor.image_processor.patch_size = {"height": 2, "width": 2}
+
+        # Test passing in an image
+        inputs_image = processor(text=prompt_string, images=[self.image_0, self.image_1], return_tensors="pt")
+        self.assertIn("input_ids", inputs_image)
+        self.assertTrue(len(inputs_image["input_ids"]) == 1)
+        self.assertIsInstance(inputs_image["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([2, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_image["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to ["USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END][IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the difference between these two images? ASSISTANT:"]
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+                    )
+        # fmt: on
+
+        # Test passing in a url
+        inputs_url = processor(text=prompt_string, images=[self.url_0, self.url_1], return_tensors="pt")
+        self.assertIn("input_ids", inputs_url)
+        self.assertTrue(len(inputs_url["input_ids"]) == 1)
+        self.assertIsInstance(inputs_url["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([2, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_url["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to ["USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END][IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the difference between these two images? ASSISTANT:"]
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+        )
+        # fmt: on
+
+        # Test passing in as a nested list
+        inputs_url = processor(text=prompt_string, images=[[self.image_0, self.image_1]], return_tensors="pt")
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([2, 3, 32, 32]))
+
+        # fmt: off
+        self.assertEqual(
+            inputs_url["input_ids"][0].tolist(),
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+        )
+        # fmt: on
+
+    def test_processor_with_multiple_images_multiple_lists(self):
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        prompt_string = [
+            "USER: [IMG][IMG]\nWhat's the difference between these two images? ASSISTANT:",
+            "USER: [IMG]\nWhat's the content of the image? ASSISTANT:",
+        ]
+        processor.tokenizer.pad_token = "</s>"
+        image_inputs = [[self.image_0, self.image_1], [self.image_2]]
+
+        # Make small for checking image token expansion
+        processor.image_processor.size = {"longest_edge": 30}
+        processor.image_processor.patch_size = {"height": 2, "width": 2}
+
+        # Test passing in an image
+        inputs_image = processor(text=prompt_string, images=image_inputs, return_tensors="pt", padding=True)
+        self.assertIn("input_ids", inputs_image)
+        self.assertTrue(len(inputs_image["input_ids"]) == 2)
+        self.assertIsInstance(inputs_image["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([3, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_image["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to ["USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END][IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the difference between these two images? ASSISTANT:"]
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+        )
+        # fmt: on
+
+        # Test passing in a url
+        inputs_url = processor(text=prompt_string, images=image_inputs, return_tensors="pt", padding=True)
+        self.assertIn("input_ids", inputs_url)
+        self.assertTrue(len(inputs_url["input_ids"]) == 2)
+        self.assertIsInstance(inputs_url["input_ids"], torch.Tensor)
+        self.assertIsInstance(inputs_image["pixel_values"], torch.Tensor)
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([3, 3, 32, 32]))
+
+        # fmt: off
+        input_ids = inputs_url["input_ids"]
+        self.assertEqual(
+            input_ids[0].tolist(),
+            # Equivalent to ["USER: [IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END][IMG][IMG][IMG_BREAK][IMG][IMG][IMG_END]\nWhat's the difference between these two images? ASSISTANT:"]
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+        )
+        # fmt: on
+
+        # Test passing as a single flat list
+        inputs_image = processor(
+            text=prompt_string, images=[self.image_0, self.image_1, self.image_2], return_tensors="pt", padding=True
+        )
+        self.assertTrue(inputs_image["pixel_values"].shape == torch.Size([3, 3, 32, 32]))
+
+        # fmt: off
+        self.assertEqual(
+            inputs_image["input_ids"][0].tolist(),
+            [21510, 1058, 1032, 10, 10, 12, 10, 10, 13, 10, 10, 12, 10, 10, 13, 1010, 7493, 1681, 1278, 6592, 2396, 2576, 2295, 8061, 1063, 1349, 4290, 16002, 41150, 1058]
+        )
+        # fmt: on
+
+    def test_processor_returns_full_length_batches(self):
+        # to avoid https://github.com/huggingface/transformers/issues/34204
+        processor = self.processor_class.from_pretrained(self.tmpdirname)
+        prompt_string = [
+            "USER: [IMG]\nWhat's the content of the image? ASSISTANT:",
+        ] * 5
+        processor.tokenizer.pad_token = "</s>"
+        image_inputs = [[self.image_0]] * 5
+
+        # Make small for checking image token expansion
+        processor.image_processor.size = {"longest_edge": 30}
+        processor.image_processor.patch_size = {"height": 2, "width": 2}
+
+        # Test passing in an image
+        inputs_image = processor(text=prompt_string, images=image_inputs, return_tensors="pt", padding=True)
+        self.assertIn("input_ids", inputs_image)
+        self.assertTrue(len(inputs_image["input_ids"]) == 5)
+        self.assertTrue(len(inputs_image["pixel_values"]) == 5)