first commit

tests/models/siglip2/test_image_processing_siglip2.py

@@ -0,0 +1,137 @@
+# 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
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class Siglip2ImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        resample=None,
+        patch_size=16,
+        max_num_patches=256,
+    ):
+        size = size if size is not None else {"height": 18, "width": 18}
+        resample = resample if resample is not None else Image.Resampling.BILINEAR
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.resample = resample
+        self.patch_size = patch_size
+        self.max_num_patches = max_num_patches
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "do_rescale": self.do_rescale,
+            "rescale_factor": self.rescale_factor,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "resample": self.resample,
+            "patch_size": self.patch_size,
+            "max_num_patches": self.max_num_patches,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.max_num_patches, self.patch_size * self.patch_size * self.num_channels
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return 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,
+        )
+
+
+@require_torch
+@require_vision
+# Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTest with CLIP->Siglip2
+class Siglip2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Siglip2ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    # Ignore copy
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "resample"))
+            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, "patch_size"))
+            self.assertTrue(hasattr(image_processing, "max_num_patches"))
+
+    # Ignore copy
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.max_num_patches, 256)
+            self.assertEqual(image_processor.patch_size, 16)
+
+            image_processor = image_processing_class.from_dict(
+                self.image_processor_dict, patch_size=32, max_num_patches=512
+            )
+            self.assertEqual(image_processor.patch_size, 32)
+            self.assertEqual(image_processor.max_num_patches, 512)
+
+    @unittest.skip(reason="not supported")
+    # Ignore copy
+    def test_call_numpy_4_channels(self):
+        pass

tests/models/siglip2/test_modeling_siglip2.py

@@ -0,0 +1,760 @@
+# 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 PyTorch Siglip2 model."""
+
+import inspect
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+from parameterized import parameterized
+from pytest import mark
+
+from transformers import Siglip2Config, Siglip2TextConfig, Siglip2VisionConfig
+from transformers.testing_utils import (
+    Expectations,
+    is_flaky,
+    require_flash_attn,
+    require_torch,
+    require_torch_accelerator,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import (
+    is_torch_available,
+    is_vision_available,
+)
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION,
+    ModelTesterMixin,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import Siglip2ForImageClassification, Siglip2Model, Siglip2TextModel, Siglip2VisionModel
+
+if is_vision_available():
+    from PIL import Image, ImageDraw
+
+    from transformers import Siglip2Processor
+
+
+class Siglip2ModelTesterMixin(ModelTesterMixin):
+    def test_sdpa_can_dispatch_composite_models(self):
+        for model_class in self.all_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+
+                # Load the model with SDPA
+                model_sdpa = model_class.from_pretrained(tmpdirname)
+
+                # Load model with eager attention
+                model_eager = model_class.from_pretrained(
+                    tmpdirname,
+                    attn_implementation="eager",
+                )
+
+            if hasattr(model_sdpa, "vision_model"):
+                self.assertTrue(model_sdpa.vision_model.config._attn_implementation == "sdpa")
+                self.assertTrue(model_eager.vision_model.config._attn_implementation == "eager")
+
+            if hasattr(model_sdpa, "text_model"):
+                self.assertTrue(model_sdpa.text_model.config._attn_implementation == "sdpa")
+                self.assertTrue(model_eager.text_model.config._attn_implementation == "eager")
+
+            self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
+            self.assertTrue(model_eager.config._attn_implementation == "eager")
+
+    @require_flash_attn
+    @require_torch_accelerator
+    @mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_inference_equivalence(self):
+        dtype = torch.float16
+
+        for model_class in self.all_model_classes:
+            if not model_class._supports_flash_attn:
+                self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
+
+            # Prepare inputs
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            if "pixel_values" in inputs_dict:
+                inputs_dict["pixel_values"] = inputs_dict["pixel_values"].to(dtype)
+
+            # Separate masks
+            attention_masks = {}
+            if "attention_mask" in inputs_dict:
+                # attention_masks["attention_mask"] = inputs_dict.pop("attention_mask")
+                inputs_dict["attention_mask"] = None
+            if "pixel_attention_mask" in inputs_dict:
+                attention_masks["pixel_attention_mask"] = inputs_dict.pop("pixel_attention_mask")
+                inputs_dict["pixel_attention_mask"] = None
+
+            # Save and load model with flash attention 2 and eager attentions
+            with tempfile.TemporaryDirectory() as tmp_dir:
+                model = model_class(config)
+                model.save_pretrained(tmp_dir)
+
+                model = model_class.from_pretrained(tmp_dir, dtype=dtype)
+                model_fa = model_class.from_pretrained(tmp_dir, dtype=dtype, attn_implementation="flash_attention_2")
+
+            model_fa.to(torch_device)
+            model.to(torch_device)
+
+            # Run forward pass without attention masks
+            with torch.no_grad():
+                outputs = model(**inputs_dict, output_hidden_states=True)
+                outputs_fa = model_fa(**inputs_dict, output_hidden_states=True)
+
+            # Choose which key to compare
+            key = [k for k in ["logits", "logits_per_image", "last_hidden_state"] if k in outputs][0]
+
+            torch.testing.assert_close(outputs[key], outputs_fa[key], atol=4e-2, rtol=4e-2)
+
+            # Run forward pass with attention masks
+            inputs_dict.update(attention_masks)
+            with torch.no_grad():
+                outputs = model(**inputs_dict, output_hidden_states=True)
+                outputs_fa = model_fa(**inputs_dict, output_hidden_states=True)
+
+            output_tensor = outputs[key]
+            output_tensor_fa = outputs_fa[key]
+
+            # Mask out padded tokens, they are different for SDPA and Flash Attention 2
+            if key == "last_hidden_state" and "pixel_attention_mask" in inputs_dict:
+                output_tensor = output_tensor * inputs_dict["pixel_attention_mask"][..., None]
+                output_tensor_fa = output_tensor_fa * inputs_dict["pixel_attention_mask"][..., None]
+            elif key == "last_hidden_state" and inputs_dict.get("attention_mask", None) is not None:
+                output_tensor = output_tensor * inputs_dict["attention_mask"][..., None]
+                output_tensor_fa = output_tensor_fa * inputs_dict["attention_mask"][..., None]
+
+            torch.testing.assert_close(output_tensor, output_tensor_fa, atol=4e-2, rtol=4e-2)
+
+            # Check with inference + dropout
+            model.train()
+            _ = model_fa(**inputs_dict, output_hidden_states=True)
+
+    @unittest.skip(reason="Siglip2 has default right padding (tested in test_flash_attn_2_inference_equivalence)")
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
+        pass
+
+    @unittest.skip(reason="SDPA can't dispatch on flash with not None `attention_mask`")
+    def test_sdpa_can_dispatch_on_flash(self):
+        pass
+
+
+class Siglip2VisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        num_patches=16,
+        image_num_patches=24,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=64,
+        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.num_patches = num_patches
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        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
+        self.seq_length = image_num_patches
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [self.batch_size, self.seq_length, self.num_channels * self.patch_size * self.patch_size]
+        )
+        pixel_attention_mask = torch.zeros(self.batch_size, self.seq_length, device=torch_device, dtype=torch.long)
+
+        spatial_shapes = [
+            (height, width)
+            for height in range(1, self.seq_length)
+            for width in range(1, self.seq_length)
+            if height * width <= self.seq_length
+        ] * self.batch_size
+        spatial_shapes = spatial_shapes[: self.batch_size]
+        spatial_shapes = torch.tensor(spatial_shapes, device=torch_device, dtype=torch.long)
+
+        for i, (height, width) in enumerate(spatial_shapes):
+            pixel_attention_mask[i, : height * width] = 1
+
+        config = self.get_config()
+
+        return config, pixel_values, pixel_attention_mask, spatial_shapes
+
+    def get_config(self):
+        return Siglip2VisionConfig(
+            num_patches=self.num_patches,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            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 create_and_check_model(self, config, pixel_values, pixel_attention_mask, spatial_shapes):
+        model = Siglip2VisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values, pixel_attention_mask, spatial_shapes)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values, pixel_attention_mask, spatial_shapes = self.prepare_config_and_inputs()
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "pixel_attention_mask": pixel_attention_mask,
+            "spatial_shapes": spatial_shapes,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Siglip2VisionModelTest(Siglip2ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as SIGLIP2 does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (Siglip2VisionModel,) if is_torch_available() else ()
+    additional_model_inputs = ["pixel_attention_mask", "spatial_shapes"]
+
+    test_resize_embeddings = False
+    # MP works but offload doesn't work when the MultiheadAttention is offloaded
+    # TODO: One potential solution would be to add to set preload_module_classes = ["Siglip2MultiheadAttentionPoolingHead"]
+    # in the dispatch_model function
+    test_cpu_offload = False
+    test_disk_offload_safetensors = False
+    test_disk_offload_bin = False
+
+    def setUp(self):
+        self.model_tester = Siglip2VisionModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=Siglip2VisionConfig, has_text_modality=False, hidden_size=32
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="SIGLIP2 does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    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(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/siglip2-base-patch16-naflex"
+        model = Siglip2VisionModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    @parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
+    @is_flaky()
+    def test_eager_matches_sdpa_inference(self, *args):
+        # adding only flaky decorator here and call the parent test method
+        return getattr(ModelTesterMixin, self._testMethodName)(self)
+
+
+class Siglip2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=64,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        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.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return Siglip2TextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            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,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = Siglip2TextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class Siglip2TextModelTest(Siglip2ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Siglip2TextModel,) if is_torch_available() else ()
+    test_resize_embeddings = False
+
+    model_split_percents = [0.5, 0.8, 0.9]
+
+    def setUp(self):
+        self.model_tester = Siglip2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Siglip2TextConfig, hidden_size=32)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="This module does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        pass
+
+    @unittest.skip(reason="Siglip2 does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/siglip2-base-patch16-naflex"
+        model = Siglip2TextModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+
+class Siglip2ModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = Siglip2TextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = Siglip2VisionModelTester(parent, **vision_kwargs)
+        self.batch_size = self.text_model_tester.batch_size  # need bs for batching_equivalence test
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values, pixel_attention_mask, spatial_shapes = (
+            self.vision_model_tester.prepare_config_and_inputs()
+        )
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values, pixel_attention_mask, spatial_shapes
+
+    def get_config(self):
+        return Siglip2Config(
+            text_config=self.text_model_tester.get_config().to_dict(),
+            vision_config=self.vision_model_tester.get_config().to_dict(),
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, attention_mask, pixel_values, pixel_attention_mask, spatial_shapes
+    ):
+        model = Siglip2Model(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, pixel_attention_mask, spatial_shapes, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values, pixel_attention_mask, spatial_shapes = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "pixel_values": pixel_values,
+            "pixel_attention_mask": pixel_attention_mask,
+            "spatial_shapes": spatial_shapes,
+            "attention_mask": attention_mask,
+            "position_ids": None,
+            "return_loss": False,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Siglip2ModelTest(Siglip2ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (Siglip2Model,) if is_torch_available() else ()
+    pipeline_model_mapping = {"feature-extraction": Siglip2Model} if is_torch_available() else {}
+    additional_model_inputs = [
+        "pixel_values",
+        "pixel_attention_mask",
+        "spatial_shapes",
+    ]
+
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    # MP works but offload doesn't work when the MultiheadAttention is offloaded
+    # TODO: One potential solution would be to add to set preload_module_classes = ["Siglip2MultiheadAttentionPoolingHead"]
+    # in the dispatch_model function
+    test_cpu_offload = False
+    test_disk_offload_safetensors = False
+    test_disk_offload_bin = False
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Siglip2ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Siglip2Config, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="Siglip2Model does not have input/output embeddings")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    def test_load_vision_text_config(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save Siglip2Config and check if we can load Siglip2VisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = Siglip2VisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save Siglip2Config and check if we can load Siglip2TextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = Siglip2TextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @unittest.skip(reason="The SigLIP2 family currently does not work with output_attentions.")
+    def test_get_text_features_attentions(self):
+        # This test should no longer be skipped once this architecture is refactored to work with output_attentions.
+        pass
+
+    @unittest.skip(reason="The SigLIP2 family currently does not work with output_hidden_states.")
+    def test_get_text_features_hidden_states(self):
+        # This test should no longer be skipped once this architecture is refactored to work with output_hidden_states.
+        pass
+
+    @unittest.skip(reason="The SigLIP2 family currently does not work with output_attentions.")
+    def test_get_image_features_attentions(self):
+        # This test should no longer be skipped once this architecture is refactored to work with output_attentions.
+        pass
+
+    @unittest.skip(reason="The SigLIP2 family currently does not work with output_hidden_states.")
+    def test_get_image_features_hidden_states(self):
+        # This test should no longer be skipped once this architecture is refactored to work with output_hidden_states.
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "google/siglip2-base-patch16-naflex"
+        model = Siglip2Model.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    @require_flash_attn
+    @require_torch_accelerator
+    @mark.flash_attn_test
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
+        self.skipTest("Siglip2 does not support right padding")
+
+
+class Siglip2ForImageClassificationModelTester(Siglip2ModelTester):
+    def __init__(self, parent):
+        super().__init__(parent)
+        self.batch_size = self.vision_model_tester.batch_size
+        self.num_hidden_layers = self.vision_model_tester.num_hidden_layers
+        self.hidden_size = self.vision_model_tester.hidden_size
+        self.seq_length = self.vision_model_tester.seq_length
+
+    def prepare_config_and_inputs(self):
+        _, pixel_values, pixel_attention_mask, spatial_shapes = self.vision_model_tester.prepare_config_and_inputs()
+        config = self.get_config()
+
+        return config, pixel_values, pixel_attention_mask, spatial_shapes
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, pixel_attention_mask, spatial_shapes = config_and_inputs
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "pixel_attention_mask": pixel_attention_mask,
+            "spatial_shapes": spatial_shapes,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Siglip2ForImageClassificationModelTest(Siglip2ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (Siglip2ForImageClassification,) if is_torch_available() else ()
+    pipeline_model_mapping = {"image-classification": Siglip2ForImageClassification} if is_torch_available() else {}
+    additional_model_inputs = ["pixel_values", "pixel_attention_mask", "spatial_shapes"]
+
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    # MP works but offload doesn't work when the MultiheadAttention is offloaded
+    # TODO: One potential solution would be to add to set preload_module_classes = ["Siglip2MultiheadAttentionPoolingHead"]
+    # in the dispatch_model function
+    test_cpu_offload = False
+    test_disk_offload_safetensors = False
+    test_disk_offload_bin = False
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = Siglip2ForImageClassificationModelTester(self)
+
+    @unittest.skip(reason="Siglip2ForImageClassification does not support inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Siglip2ForImageClassification does not support inputs_embeds")
+    def test_model_get_set_embeddings(self):
+        pass
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing(self):
+        super().test_training_gradient_checkpointing()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        super().test_training_gradient_checkpointing_use_reentrant_false()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        super().test_training_gradient_checkpointing_use_reentrant_true()
+
+
+# Draw a circle on an images with different aspect ratios
+def prepare_images():
+    shapes = [(224, 224), (1024, 1024), (224, 1024)]
+    images = []
+    for height, width in shapes:
+        image = Image.new("RGB", (width, height), color="red")
+        draw = ImageDraw.Draw(image)
+        center_x = image.width // 2
+        center_y = image.height // 2
+        radius = min(center_x, center_y) // 8 * 7
+        draw.ellipse(
+            (center_x - radius, center_y - radius, center_x + radius, center_y + radius),
+            fill="blue",
+            outline="green",
+            width=image.width // 20,
+        )
+        images.append(image)
+    return images
+
+
+@require_vision
+@require_torch
+class Siglip2ModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        model_name = "google/siglip2-base-patch16-naflex"
+        model = Siglip2Model.from_pretrained(model_name).to(torch_device)
+        processor = Siglip2Processor.from_pretrained(model_name)
+
+        images = prepare_images()
+        text = [
+            "circle",
+            "ellipsoid",
+            "blue circle on red background",
+            "blue circle with green border on red background",
+            "green circle on red background",
+            "a dog",
+            "a blue dog with a green border on a red background",
+        ]
+
+        inputs = processor(text=text, images=images, return_tensors="pt")
+        inputs = inputs.to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        logits_per_image = outputs.logits_per_image
+        logits_per_text = outputs.logits_per_text
+
+        # verify the logits shape
+        self.assertEqual(
+            logits_per_image.shape,
+            torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
+        )
+        self.assertEqual(
+            logits_per_text.shape,
+            torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
+        )
+
+        # verify the logits values
+        # fmt: off
+        expected_logits_per_texts = Expectations({
+            ("cuda", None): [
+                [  1.0195,  -0.0280,  -1.4468], [ -4.5395,  -6.2269,  -1.5667], [  4.1757,   5.0358,   3.5159],
+                [  9.4264,  10.1879,   6.3353], [  2.4409,   3.1058,   4.5491], [-12.3230, -13.7355, -13.4632],
+                [  1.1520,   1.1687,  -1.9647],
+            ],
+            ("rocm", (9, 5)): [
+                [  1.0236,  -0.0376,  -1.4464], [ -4.5358,  -6.2235,  -1.5628], [  4.1708,   5.0334,   3.5187],
+                [  9.4241,  10.1828,   6.3366], [  2.4371,   3.1062,   4.5530], [-12.3173, -13.7240, -13.4580],
+                [  1.1502,   1.1716,  -1.9623]
+            ],
+            ("xpu", 3): [
+                [  1.0195,  -0.0280,  -1.4468], [ -4.5395,  -6.2269,  -1.5667], [  4.1757,   5.0358,   3.5159],
+                [  9.4264,  10.1879,   6.3353], [  2.4409,   3.1058,   4.5491], [-12.3230, -13.7355, -13.4632],
+                [  1.1520,   1.1687,  -1.9647]
+            ],
+        })
+        EXPECTED_LOGITS_PER_TEXT = torch.tensor(expected_logits_per_texts.get_expectation()).to(torch_device)
+        # fmt: on
+
+        torch.testing.assert_close(outputs.logits_per_text, EXPECTED_LOGITS_PER_TEXT, rtol=1e-3, atol=1e-3)

tests/models/siglip2/test_tokenization_siglip2.py

@@ -0,0 +1,75 @@
+# 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.
+
+import tempfile
+import unittest
+
+from transformers import Siglip2Tokenizer
+from transformers.testing_utils import require_tokenizers
+
+
+@require_tokenizers
+class Siglip2TokenizerTest(unittest.TestCase):
+    """
+    Integration test for Siglip2Tokenizer:
+    - verify hub loading,
+    - default lowercasing behavior,
+    - save/load roundtrip.
+    """
+
+    from_pretrained_id = "google/siglip2-base-patch16-224"
+
+    def test_tokenizer(self):
+        tokenizer = Siglip2Tokenizer.from_pretrained(self.from_pretrained_id)
+
+        texts_uc = [
+            "HELLO WORLD!",
+            "Hello   World!!",
+            "A Picture Of Zürich",
+            "San Francisco",
+            "MIXED-case: TeSt 123",
+        ]
+        texts_lc = [t.lower() for t in texts_uc]
+
+        # default lowercasing (single + batch paths)
+        for t_uc, t_lc in zip(texts_uc, texts_lc):
+            with self.subTest(text=t_uc):
+                enc_uc = tokenizer(t_uc, truncation=True)
+                enc_lc = tokenizer(t_lc, truncation=True)
+                self.assertListEqual(enc_uc["input_ids"], enc_lc["input_ids"])
+
+        batch_uc = tokenizer(texts_uc, truncation=True)
+        batch_lc = tokenizer(texts_lc, truncation=True)
+        self.assertListEqual(batch_uc["input_ids"], batch_lc["input_ids"])
+
+        # padding/truncation path (avoid relying on model_max_length)
+        max_len = 64
+        padded = tokenizer(texts_uc, padding="max_length", truncation=True, max_length=max_len)
+        # ensure every sequence is padded/truncated to max_len
+        for seq in padded["input_ids"]:
+            self.assertEqual(len(seq), max_len)
+
+        # save/load roundtrip preserves behavior
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tokenizer.save_pretrained(tmpdir)
+            tokenizer_reloaded = Siglip2Tokenizer.from_pretrained(tmpdir)
+
+            batch_uc_2 = tokenizer_reloaded(texts_uc, truncation=True)
+            batch_lc_2 = tokenizer_reloaded(texts_lc, truncation=True)
+            self.assertListEqual(batch_uc_2["input_ids"], batch_lc_2["input_ids"])
+            self.assertListEqual(batch_uc["input_ids"], batch_uc_2["input_ids"])
+
+            padded_2 = tokenizer_reloaded(texts_uc, padding="max_length", truncation=True, max_length=max_len)
+            for seq in padded_2["input_ids"]:
+                self.assertEqual(len(seq), max_len)