[EXPERIMENT][WIP][OpenVINO] Add support for DINOv3 ConvNeXt

docs/source/openvino/models.mdx

@@ -51,6 +51,7 @@ Here is the list of the supported architectures :
 - DeepSeek
 - DeepSeek-V2
 - DeepSeek-V3
+- DINOv3 (ConvNeXt variant)
 - DistilBERT
 - ERNIE 4.5
 - ELECTRA

optimum/exporters/openvino/model_configs.py

@@ -218,7 +218,6 @@
     NormalizedVisionConfig,
 )
 
-
 COMMON_TEXT_TASKS = [
     "feature-extraction",
     "fill-mask",
@@ -5842,6 +5841,11 @@ class ConvNextOpenVINOConfig(ViTOpenVINOConfig):
     pass
 
 
+@register_in_tasks_manager("dinov3_convnext", *["feature-extraction"])
+class DINOv3ConvNextOpenVINOConfig(ViTOpenVINOConfig):
+    pass
+
+
 @register_in_tasks_manager("resnet", *["feature-extraction", "image-classification"])
 class ResNetOpenVINOConfig(ViTOpenVINOConfig):
     pass

tests/openvino/test_modeling.py

@@ -117,7 +117,6 @@
 )
 from optimum.utils.testing_utils import require_diffusers
 
-
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
 
@@ -1545,6 +1544,9 @@ def test_compare_to_transformers(self, model_arch):
 class OVModelForCustomTasksIntegrationTest(unittest.TestCase):
     SUPPORTED_ARCHITECTURES_WITH_ATTENTION = ["vit-with-attentions"]
     SUPPORTED_ARCHITECTURES_WITH_HIDDEN_STATES = ["vit-with-hidden-states"]
+    # Vision feature-extraction backbones (pixel_values -> last_hidden_state, no classification
+    # head), unlike OVModelForFeatureExtraction which is hard-coded for text inputs.
+    SUPPORTED_ARCHITECTURES_VISION_FEATURE_EXTRACTION = ["dinov3_convnext"]
 
     def _get_sample_image(self):
         url = TEST_IMAGE_URL
@@ -1627,6 +1629,39 @@ def test_compare_output_hidden_states(self, model_arch):
         del ov_model
         gc.collect()
 
+    @parameterized.expand(SUPPORTED_ARCHITECTURES_VISION_FEATURE_EXTRACTION)
+    def test_compare_output_vision_feature_extraction(self, model_arch):
+        model_id = MODEL_NAMES[model_arch]
+
+        image = self._get_sample_image()
+        preprocessor = AutoImageProcessor.from_pretrained(model_id)
+        inputs = preprocessor(images=image, return_tensors="pt")
+
+        transformers_model = AutoModel.from_pretrained(model_id)
+        transformers_model.eval()
+        with torch.no_grad():
+            transformers_outputs = transformers_model(**inputs)
+
+        ov_model = OVModelForCustomTasks.from_pretrained(
+            model_id, export=True, task="feature-extraction", ov_config=F32_CONFIG, device=OPENVINO_DEVICE
+        )
+        self.assertIsInstance(ov_model.config, PretrainedConfig)
+
+        for input_type in ["pt", "np"]:
+            inputs = preprocessor(images=image, return_tensors=input_type)
+            ov_outputs = ov_model(**inputs)
+            self.assertIn("last_hidden_state", ov_outputs)
+            self.assertIsInstance(ov_outputs.last_hidden_state, TENSOR_ALIAS_TO_TYPE[input_type])
+            self.assertTrue(
+                torch.allclose(
+                    torch.Tensor(ov_outputs.last_hidden_state), transformers_outputs.last_hidden_state, atol=1e-4
+                )
+            )
+
+        del transformers_model
+        del ov_model
+        gc.collect()
+
 
 class OVModelForOpenCLIPZeroShortImageClassificationTest(unittest.TestCase):
     OV_MODEL_ID = HUB_MODEL_NAMES["open-clip"]

tests/openvino/utils_tests.py

@@ -192,6 +192,7 @@ def _create_tiny_kokoro_model():
     "deit": "optimum-intel-internal-testing/tiny-random-DeiTModel",
     "convnext": "optimum-intel-internal-testing/tiny-random-convnext",
     "convnextv2": "optimum-intel-internal-testing/tiny-random-ConvNextV2Model",
+    "dinov3_convnext": "optimum-intel-internal-testing/tiny-random-dinov3_convnext",
     "distilbert": "optimum-intel-internal-testing/tiny-random-distilbert",
     "distilbert-ov": "optimum-intel-internal-testing/ov-tiny-random-distilbert",
     "donut": "optimum-internal-testing/tiny-random-VisionEncoderDecoderModel-donut",