init

transformers/tests/models/lfm2_vl/test_image_processing_lfm2_vl.py

@@ -0,0 +1,289 @@
+# coding=utf-8
+# Copyright 2025 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
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_torch_available():
+    import torch
+
+    if is_torchvision_available():
+        from transformers import Lfm2VlImageProcessorFast
+        from transformers.models.lfm2_vl.image_processing_lfm2_vl_fast import find_closest_aspect_ratio
+
+
+class Lfm2VlImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_images=1,
+        min_resolution=256,
+        max_resolution=1024,
+        downsample_factor=2,
+        do_image_splitting=False,
+        min_tiles=2,
+        max_tiles=10,
+        use_thumbnail=True,
+        min_image_tokens=64,
+        max_image_tokens=256,
+        encoder_patch_size=16,
+        tile_size=512,
+        max_pixels_tolerance=2.0,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.num_images = num_images
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+
+        self.downsample_factor = downsample_factor
+        self.do_image_splitting = do_image_splitting
+        self.min_tiles = min_tiles
+        self.max_tiles = max_tiles
+        self.use_thumbnail = use_thumbnail
+        self.min_image_tokens = min_image_tokens
+        self.max_image_tokens = max_image_tokens
+        self.encoder_patch_size = encoder_patch_size
+        self.tile_size = tile_size
+        self.max_pixels_tolerance = max_pixels_tolerance
+
+    def prepare_image_processor_dict(self):
+        return {
+            "downsample_factor": self.downsample_factor,
+            "do_image_splitting": self.do_image_splitting,
+            "min_tiles": self.min_tiles,
+            "max_tiles": self.max_tiles,
+            "use_thumbnail": self.use_thumbnail,
+            "min_image_tokens": self.min_image_tokens,
+            "max_image_tokens": self.max_image_tokens,
+            "encoder_patch_size": self.encoder_patch_size,
+            "tile_size": self.tile_size,
+            "max_pixels_tolerance": self.max_pixels_tolerance,
+        }
+
+    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 [[image] for image in images]
+
+
+@require_torch
+@require_vision
+class Lfm2VlImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    test_slow_image_processor = False
+    fast_image_processing_class = Lfm2VlImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Lfm2VlImageProcessingTester(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, "downsample_factor"))
+            self.assertTrue(hasattr(image_processing, "min_tiles"))
+            self.assertTrue(hasattr(image_processing, "max_tiles"))
+            self.assertTrue(hasattr(image_processing, "use_thumbnail"))
+            self.assertTrue(hasattr(image_processing, "min_image_tokens"))
+            self.assertTrue(hasattr(image_processing, "max_image_tokens"))
+            self.assertTrue(hasattr(image_processing, "encoder_patch_size"))
+            self.assertTrue(hasattr(image_processing, "tile_size"))
+            self.assertTrue(hasattr(image_processing, "max_pixels_tolerance"))
+
+    @require_vision
+    def test_smart_resize(self):
+        # verify that smart resize output dims are divisible by encoder_patch_size * downsample_factor
+        image_processing = self.fast_image_processing_class(**self.image_processor_dict)
+        width, height = image_processing.smart_resize(
+            height=500,
+            width=300,
+            downsample_factor=image_processing.downsample_factor,
+            min_image_tokens=image_processing.min_image_tokens,
+            max_image_tokens=image_processing.max_image_tokens,
+            encoder_patch_size=image_processing.encoder_patch_size,
+        )
+        mod = image_processing.encoder_patch_size * image_processing.downsample_factor
+        self.assertEqual(width % mod, 0)
+        self.assertEqual(height % mod, 0)
+
+    @require_vision
+    def test_get_grid_layout(self):
+        # splitting a 512×512 image into tiles of size processor.image_processor.tile_size
+        image_processing = self.fast_image_processing_class(**self.image_processor_dict)
+        rows, cols, _, _, num_patches = image_processing._get_grid_layout(
+            height=1024,
+            width=1024,
+            min_tiles=image_processing.min_tiles,
+            max_tiles=image_processing.max_tiles,
+            tile_size=image_processing.tile_size,
+        )
+        self.assertEqual(num_patches, 4)
+        self.assertEqual(num_patches, rows * cols)
+
+        rows, cols, _, _, num_patches = image_processing._get_grid_layout(
+            height=1024,
+            width=1024,
+            min_tiles=8,
+            max_tiles=8,
+            tile_size=image_processing.tile_size,
+        )
+        self.assertEqual(num_patches, 8)
+        self.assertEqual(num_patches, rows * cols)
+
+    def test_find_closest_aspect_ratio(self):
+        # should pick (1,1) over (2,1) for a square image
+        result = find_closest_aspect_ratio(1.0, [(1, 1), (2, 1)], width=100, height=100, image_size=100)
+        self.assertEqual(result, (1, 1))
+
+        result = find_closest_aspect_ratio(0.5, [(1, 1), (1, 2)], width=100, height=200, image_size=200)
+        self.assertEqual(result, (1, 2))
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        image_processing = self.fast_image_processing_class(**self.image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+        for sample_images in image_inputs:
+            for image in sample_images:
+                self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_numpy_4_channels(self):
+        # Lfm2Vl always processes images as RGB, so it always returns images with 3 channels
+        # Initialize image_processing
+        image_processor_dict = self.image_processor_dict
+        image_processing = self.fast_image_processing_class(**image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+
+        for sample_images in image_inputs:
+            for image in sample_images:
+                self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        image_processing = self.fast_image_processing_class(**self.image_processor_dict)
+        # create random PIL images
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        image_processing = self.fast_image_processing_class(**self.image_processor_dict)
+        # create random PyTorch tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )

transformers/tests/models/lfm2_vl/test_modeling_lfm2_vl.py

@@ -0,0 +1,296 @@
+# Copyright 2025 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 LFM2-VL model."""
+
+import math
+import unittest
+from io import BytesIO
+
+import pytest
+import requests
+
+from transformers import AutoProcessor, is_torch_available
+from transformers.models.lfm2_vl.modeling_lfm2_vl import Lfm2VlForConditionalGeneration
+from transformers.testing_utils import (
+    cleanup,
+    require_read_token,
+    require_torch,
+    require_torch_accelerator,
+    slow,
+    torch_device,
+)
+from transformers.utils.import_utils import is_vision_available
+
+from ...causal_lm_tester import CausalLMModelTester
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_vision_available():
+    from PIL import Image
+
+if is_torch_available():
+    import torch
+
+    from transformers import Lfm2VlConfig, Lfm2VlForConditionalGeneration, Lfm2VlModel
+
+
+class Lfm2VlModelTester(CausalLMModelTester):
+    if is_torch_available():
+        config_class = Lfm2VlConfig
+        base_model_class = Lfm2VlModel
+        causal_lm_class = Lfm2VlForConditionalGeneration
+
+    def __init__(
+        self,
+        parent,
+        is_training=True,
+        batch_size=2,
+        scale_factor=2,
+        num_images=2,
+        vision_config={
+            "hidden_size": 32,
+            "intermediate_size": 37,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 2,
+            "num_channels": 3,
+            "num_patches": 16,
+            "patch_size": 4,
+            "hidden_act": "gelu_pytorch_tanh",
+            "layer_norm_eps": 1e-6,
+            "attention_dropout": 0.0,
+        },
+        text_config={
+            "vocab_size": 100,
+            "hidden_size": 32,
+            "intermediate_size": 37,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "num_key_value_heads": 2,
+            "max_position_embeddings": 100,
+            "pad_token_id": 0,
+            "bos_token_id": 1,
+            "eos_token_id": 2,
+            "tie_word_embeddings": True,
+            "rope_theta": 1000000.0,
+            "conv_bias": False,
+            "conv_L_cache": 3,
+            "block_multiple_of": 2,
+            "full_attn_idxs": [0],
+        },
+        image_token_id=4,
+        downsample_factor=4,
+        projector_hidden_size=32,
+    ):
+        super().__init__(parent)
+        self.vision_config = vision_config
+        self.text_config = text_config
+        self.image_token_id = image_token_id
+        self.is_training = is_training
+        self.batch_size = batch_size
+        self.scale_factor = scale_factor
+        self.num_images = num_images
+        self.downsample_factor = downsample_factor
+        self.projector_hidden_size = projector_hidden_size
+        self.image_seq_length = 4
+
+    def get_config(self):
+        return Lfm2VlConfig(
+            vision_config=self.vision_config,
+            text_config=self.text_config,
+            image_token_id=self.image_token_id,
+            downsample_factor=self.downsample_factor,
+            projector_hidden_size=self.projector_hidden_size,
+        )
+
+    def prepare_config_and_inputs(self):
+        # Create dummy pixel values: [num_images, num_patches, channels * patch_size^2]
+        patch_size = self.vision_config["patch_size"]
+        pixel_values = floats_tensor([self.num_images, 64, 3 * patch_size * patch_size])
+
+        # Spatial shapes: one (height_patches, width_patches) per image
+        patches = int(math.sqrt(64))
+        spatial_shapes = torch.tensor([[patches, patches]] * self.num_images, dtype=torch.long, device=torch_device)
+
+        # Pixel attention mask: mark all patches as valid (no padding)
+        pixel_attention_mask = torch.ones((self.num_images, 64), dtype=torch.long, device=torch_device)
+        config = self.get_config()
+        return config, pixel_values, spatial_shapes, pixel_attention_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, spatial_shapes, pixel_attention_mask = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 2) + 1
+
+        # For simplicity just set the last n tokens to the image token
+        input_ids[input_ids == self.image_token_id] = self.text_config["pad_token_id"]
+        input_ids[:, -self.image_seq_length :] = self.image_token_id
+
+        attention_mask = input_ids.ne(1).to(torch_device)
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "spatial_shapes": spatial_shapes,
+            "pixel_attention_mask": pixel_attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Lfm2VlModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (Lfm2VlModel, Lfm2VlForConditionalGeneration) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": Lfm2VlModel,
+            "text-generation": Lfm2VlForConditionalGeneration,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+    fx_compatible = False
+    model_tester_class = Lfm2VlModelTester
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Lfm2VlModelTester(self)
+        common_properties = ["image_token_id", "projector_hidden_size"]
+        self.config_tester = ConfigTester(
+            self, config_class=Lfm2VlConfig, has_text_modality=False, common_properties=common_properties
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(
+        "Lfm2 backbone alternates between attention and conv layers, so attention are only returned for attention layers"
+    )
+    def test_attention_outputs(self):
+        pass
+
+    @unittest.skip("Lfm2 backbone has a special cache format as it alternates between attention and conv layers")
+    def test_past_key_values_format(self):
+        pass
+
+    @unittest.skip(
+        "Lfm2 backbone has a special cache format which is not compatible with compile as it has static address for conv cache"
+    )
+    @pytest.mark.torch_compile_test
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+    @unittest.skip(reason="Backbone Siglip2VisionModel does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="Backbone Siglip2VisionModel does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        pass
+
+    @unittest.skip(reason="Backbone Siglip2VisionModel does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(
+        reason="Siglip2 backbone has a non-standard initialization scheme, that this test cannot handle easily"
+    )
+    def test_initialization(self):
+        pass
+
+
+@require_torch_accelerator
+@require_read_token
+@slow
+class Lfm2VlForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("LiquidAI/LFM2-VL-1.6B")
+        self.processor.tokenizer.padding_side = "left"
+        self.image = Image.open(
+            requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw
+        )
+        self.image2 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def test_integration_test(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = "In this image, we see a cat and a dog lying on a pink blanket. They are both sleeping peacefully. They are"
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    def test_integration_test_high_resolution(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image2
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = (
+            "In this image, we see the Statue of Liberty, standing tall on its pedestal. The statue is made of metal,"
+        )
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    def test_integration_test_batched(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-450M",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = ["<image>In this image, we see", "<image>In this image, there is a cat on"]
+        images = [[self.image2], [self.image]]
+        inputs = self.processor(text=text, images=images, return_tensors="pt", padding=True)
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = [
+            "In this image, we see a panoramic view of the New York City skyline. The iconic Statics and the New York",
+            "In this image, there is a cat on a bed with a cat on a bed with a cat on a bed with a cat on a bed",
+        ]
+        self.assertListEqual(generated_texts, expected_generated_text)

transformers/tests/models/lfm2_vl/test_processing_lfm2_vl.py

@@ -0,0 +1,467 @@
+# Copyright 2025 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 math
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers import AutoTokenizer, Lfm2VlProcessor
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torchvision_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from PIL import Image
+
+    if is_torchvision_available():
+        from transformers import Lfm2VlImageProcessorFast
+
+
+@require_torch
+@require_vision
+class Lfm2VlProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Lfm2VlProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+        processor_kwargs = cls.prepare_processor_dict()
+        image_processor = Lfm2VlImageProcessorFast(
+            tile_size=14,
+            min_image_tokens=2,
+            max_image_tokens=10,
+            encoder_patch_size=2,
+            do_image_splitting=False,
+        )
+        tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM2-VL-1.6B", **processor_kwargs)
+
+        processor = Lfm2VlProcessor(tokenizer=tokenizer, image_processor=image_processor, **processor_kwargs)
+        processor.save_pretrained(cls.tmpdirname)
+
+        # Create images with different sizes
+        cls.small_image = Image.new("RGB", (256, 256))
+        cls.large_image = Image.new("RGB", (512, 1024))
+        cls.high_res_image = Image.new("RGB", (1024, 1024))
+
+        cls.bos_token = processor.tokenizer.bos_token
+        cls.image_token = processor.image_token
+
+        cls.bos_token_id = processor.tokenizer.convert_tokens_to_ids(cls.bos_token)
+        cls.image_token_id = processor.image_token_id
+        cls.image_start_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_start_token)
+        cls.image_end_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_end_token)
+        cls.padding_token_id = processor.tokenizer.pad_token_id
+        cls.image_thumbnail_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_thumbnail_token)
+
+    def get_tokenizer(self, **kwargs):
+        return Lfm2VlProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return Lfm2VlProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def get_processor(self, **kwargs):
+        return Lfm2VlProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    @staticmethod
+    def prepare_processor_dict():
+        chat_template = (
+            "{{bos_token}}{% for message in messages %}"
+            "{{'<|im_start|>' + message['role'] + '\n'}}"
+            "{% if message['content'] is string %}"
+            "{{ message['content'] }}"
+            "{% else %}"
+            "{% for content in message['content'] %}"
+            "{% if content['type'] == 'image' %}"
+            "{{ '<image>' }}"
+            "{% elif content['type'] == 'text' %}"
+            "{{ content['text'] }}"
+            "{% endif %}"
+            "{% endfor %}"
+            "{% endif %}"
+            "{{'<|im_end|>\n'}}"
+            "{% endfor %}"
+            "{% if add_generation_prompt %}"
+            "{{'<|im_start|>assistant\n' }}"
+            "{% endif %}"
+        )
+        return {"chat_template": chat_template, "use_image_special_tokens": True}
+
+    # Override as Lfm2VL needs images/video to be an explicitly nested batch
+    def prepare_image_inputs(self, batch_size=None):
+        """This function prepares a list of PIL images for testing"""
+        images = super().prepare_image_inputs(batch_size)
+        if isinstance(images, (list, tuple)):
+            images = [[image] for image in images]
+        return images
+
+    def get_split_image_expected_tokens(self, processor, image_rows, image_cols, add_thumbnail, image_seq_len):
+        text_split_images = [self.image_start_token_id]
+        num_patches_tile = processor.image_processor.tile_size // processor.image_processor.encoder_patch_size
+        tile_seq_len = math.ceil(num_patches_tile / processor.image_processor.downsample_factor) ** 2
+        for n_h in range(image_rows):
+            for n_w in range(image_cols):
+                text_split_images += (
+                    processor.tokenizer(f"<|img_row_{n_h + 1}_col_{n_w + 1}|>", add_special_tokens=False)["input_ids"]
+                    + [self.image_token_id] * tile_seq_len
+                )
+        if add_thumbnail:
+            text_split_images += [self.image_thumbnail_token_id] + [self.image_token_id] * image_seq_len
+        text_split_images += [self.image_end_token_id]
+        return text_split_images
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    def test_process_interleaved_images_prompts_no_image_splitting_single_image(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+        image_str = "<image>"
+
+        # Test that a single image is processed correctly
+        inputs = processor(images=self.small_image, text=image_str)
+        encoder_feature_dims = (
+            3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_values"]).shape,
+            (1, processor.image_processor.max_num_patches, encoder_feature_dims),
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_attention_mask"]).shape, (1, processor.image_processor.max_num_patches)
+        )
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6]])
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_no_image_splitting_single_image_with_text(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+        inputs = processor(text=text, images=self.small_image)
+
+        # fmt: off
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        expected_input_ids = [[self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id] + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        encoder_feature_dims = 3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        self.assertEqual(np.array(inputs["pixel_values"]).shape, (1, processor.image_processor.max_num_patches, encoder_feature_dims))
+        self.assertEqual(np.array(inputs["pixel_attention_mask"]).shape, (1, processor.image_processor.max_num_patches))
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6]])
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_no_image_splitting_multiple_images(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "In this image, we see"
+
+        text = [
+            image_str + text_str_1,
+            image_str + image_str + text_str_2,
+        ]
+        images = [[self.small_image], [self.small_image, self.small_image]]
+
+        inputs = processor(text=text, images=images, padding=True)
+
+        tokenized_sentence_1 = processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = processor.tokenizer(text_str_2, add_special_tokens=False)
+        image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+        expected_input_ids_1 = image_tokens + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = 2 * image_tokens + tokenized_sentence_2["input_ids"]
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [self.padding_token_id] * pad_len + expected_input_ids_1
+
+        self.assertEqual(inputs["input_ids"], [padded_expected_input_ids_1, expected_input_ids_2])
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)],
+        )
+        encoder_feature_dims = (
+            3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_values"]).shape,
+            (3, processor.image_processor.max_num_patches, encoder_feature_dims),
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_attention_mask"]).shape, (3, processor.image_processor.max_num_patches)
+        )
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6], [6, 6], [6, 6]])
+
+    def test_process_interleaved_images_prompts_image_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "bla, bla"
+
+        text = [image_str + text_str_1, text_str_2 + image_str + image_str]
+        images = [[self.small_image], [self.high_res_image, self.high_res_image]]
+
+        inputs = processor(
+            text=text,
+            images=images,
+            padding=True,
+            padding_side="left",
+            max_pixels_tolerance=2.0,
+            use_thumbnail=True,
+            do_image_splitting=True,
+        )
+
+        tokenized_sentence_1 = processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = processor.tokenizer(text_str_2, add_special_tokens=False)
+
+        small_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        large_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        high_res_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+
+        expected_input_ids_1 = small_image_tokens + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = tokenized_sentence_2["input_ids"] + large_image_tokens + high_res_image_tokens
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [self.padding_token_id] * pad_len + expected_input_ids_1
+
+        self.assertEqual(inputs["input_ids"][0], padded_expected_input_ids_1)
+        self.assertEqual(inputs["input_ids"][1], expected_input_ids_2)
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)],
+        )
+        self.assertEqual(np.array(inputs["pixel_values"]).shape, (30, 49, 12))
+        self.assertEqual(np.array(inputs["pixel_attention_mask"]).shape, (30, 49))
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), ([[7, 7]] * 9 + [[6, 6]]) * 3)
+
+    def test_add_special_tokens_processor_image_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = text_str + image_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.high_res_image, add_special_tokens=False, do_image_splitting=True)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        split_high_res_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        expected_input_ids = [tokenized_sentence["input_ids"] + split_high_res_image_tokens]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_add_special_tokens_processor_image_splitting_large_image(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = text_str + image_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.large_image, add_special_tokens=False, max_pixels_tolerance=2.0, do_image_splitting=True)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        large_image_tokens = self.get_split_image_expected_tokens(processor, 2, 4, True, 8)
+        expected_input_ids = [tokenized_sentence["input_ids"] + large_image_tokens]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_add_special_tokens_processor_image_no_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.high_res_image, add_special_tokens=False, use_image_special_tokens=True, do_image_splitting=False)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        split_high_res_image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+        expected_input_ids = [split_high_res_image_tokens + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_image_error(self):
+        processor = self.get_processor()
+
+        text = [
+            "This is a test sentence.",
+            "In this other sentence we try some good things",
+        ]
+        images = [[self.small_image], [self.large_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [[self.small_image], []]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        text = [
+            "This is a test sentence.<image>",
+            "In this other sentence we try some good things<image>",
+        ]
+        images = [[self.small_image], [self.large_image, self.high_res_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [[], [self.large_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [self.small_image, self.large_image, self.high_res_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [self.small_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        text = [
+            "This is a test sentence.",
+            "In this other sentence we try some good things<image>",
+        ]
+        images = [[self.small_image], []]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        images = [[], [self.large_image]]
+        processor(text=text, images=images, padding=True)
+
+        images = [self.small_image, self.large_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        images = [self.small_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+    def test_apply_chat_template(self):
+        # Message contains content which a mix of lists with images and image urls and string
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What do these images show?"},
+                    {"type": "image"},
+                    {"type": "image"},
+                ],
+            },
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": "The first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.",
+                    }
+                ],
+            },
+            {"role": "user", "content": [{"type": "text", "text": "And who is that?"}]},
+        ]
+        processor = self.get_processor()
+        # Make short sequence length to test that the fake tokens are added correctly
+        rendered = processor.apply_chat_template(messages, add_generation_prompt=True)
+
+        expected_rendered = (
+            "<|startoftext|><|im_start|>user\nWhat do these images show?<image><image><|im_end|>\n"
+            "<|im_start|>assistant\nThe first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.<|im_end|>\n"
+            "<|im_start|>user\nAnd who is that?<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        self.assertEqual(rendered, expected_rendered)
+
+    def test_text_only_inference(self):
+        """Test that the processor works correctly with text-only input."""
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        text = "This is a simple text without images."
+        inputs = processor(text=text)
+
+        tokenized_sentence = processor.tokenizer(text, add_special_tokens=False)
+        expected_input_ids = [tokenized_sentence["input_ids"]]
+
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        self.assertTrue("pixel_values" not in inputs)
+        self.assertTrue("pixel_attention_mask" not in inputs)
+
+        # Test batch of texts without image tokens
+        texts = ["First text.", "Second piece of text."]
+        batch_inputs = processor(text=texts, padding=True)
+
+        tokenized_1 = processor.tokenizer(texts[0], add_special_tokens=False)
+        tokenized_2 = processor.tokenizer(texts[1], add_special_tokens=False)
+
+        expected_1 = tokenized_1["input_ids"]
+        expected_2 = tokenized_2["input_ids"]
+
+        # Pad the shorter sequence
+        pad_len = len(expected_2) - len(expected_1)
+        if pad_len > 0:
+            padded_expected_1 = [self.padding_token_id] * pad_len + expected_1
+            expected_attention_1 = [0] * pad_len + [1] * len(expected_1)
+            self.assertEqual(batch_inputs["input_ids"], [padded_expected_1, expected_2])
+            self.assertEqual(batch_inputs["attention_mask"], [expected_attention_1, [1] * len(expected_2)])
+        else:
+            pad_len = -pad_len
+            padded_expected_2 = [self.padding_token_id] * pad_len + expected_2
+            expected_attention_2 = [0] * pad_len + [1] * len(expected_2)
+            self.assertEqual(batch_inputs["input_ids"], [expected_1, padded_expected_2])
+            self.assertEqual(batch_inputs["attention_mask"], [[1] * len(expected_1), expected_attention_2])
+
+    def test_missing_images_error(self):
+        """Test that appropriate error is raised when images are referenced but not provided."""
+        processor = self.get_processor()
+
+        # Test single text with image token but no image
+        text = "Let me show you this image: <image> What do you think?"
+        with self.assertRaises(ValueError) as context:
+            processor(text=text)
+        self.assertTrue("We detected 1 tokens in the text but no images were passed" in str(context.exception))
+
+        # Test batch with image tokens but no images
+        texts = [
+            "First text with <image> token.",
+            "Second text <image> with token.",
+        ]
+        with self.assertRaises(ValueError) as context:
+            processor(text=texts)
+        self.assertTrue("We detected 2 tokens in the text but no images were passed" in str(context.exception))
+
+        # Test with None as Images
+        with self.assertRaises(ValueError) as context:
+            processor(text=text, images=None)
+        self.assertTrue("We detected 1 tokens in the text but no images were passed" in str(context.exception))
+
+        with self.assertRaises(ValueError) as context:
+            processor(text=texts, images=None)
+        self.assertTrue("We detected 2 tokens in the text but no images were passed" in str(context.exception))