first commit

tests/models/videomt/test_modeling_videomt.py

@@ -0,0 +1,562 @@
+# Copyright 2026 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.
+"""Testing suite for the PyTorch VidEoMT model."""
+
+import unittest
+
+import numpy as np
+
+from transformers import VideomtConfig, VideomtForUniversalSegmentation
+from transformers.testing_utils import (
+    Expectations,
+    require_torch,
+    require_torch_gpu,
+    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 ModelTesterMixin, floats_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoVideoProcessor
+
+
+class VideomtForUniversalSegmentationTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        num_frames=1,
+        image_size=40,
+        patch_size=2,
+        num_queries=5,
+        num_register_tokens=19,
+        num_labels=4,
+        hidden_size=8,
+        num_attention_heads=2,
+        num_hidden_layers=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_frames = num_frames
+        self.num_queries = num_queries
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_labels = num_labels
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.num_register_tokens = num_register_tokens
+        self.is_training = False
+
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1 + self.num_register_tokens
+
+    def get_config(self):
+        config = {
+            "image_size": self.image_size,
+            "patch_size": self.patch_size,
+            "num_labels": self.num_labels,
+            "hidden_size": self.hidden_size,
+            "num_attention_heads": self.num_attention_heads,
+            "num_hidden_layers": self.num_hidden_layers,
+            "num_register_tokens": self.num_register_tokens,
+            "num_queries": self.num_queries,
+            "num_blocks": 1,
+            "rope_parameters": {"rope_theta": 100.0},
+        }
+        return VideomtConfig(**config)
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_frames, 3, self.image_size, self.image_size]).to(
+            torch_device
+        )
+
+        config = self.get_config()
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config, pixel_values = self.prepare_config_and_inputs()
+        inputs_dict = {"pixel_values_videos": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class VideomtForUniversalSegmentationTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (VideomtForUniversalSegmentation,) if is_torch_available() else ()
+    pipeline_model_mapping = {}
+    is_encoder_decoder = False
+    test_missing_keys = False
+    test_torch_exportable = False
+
+    def setUp(self):
+        self.model_tester = VideomtForUniversalSegmentationTester(self)
+        self.config_tester = ConfigTester(self, config_class=VideomtConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="VideoMT 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)
+            output_embeddings = model.get_output_embeddings()
+            self.assertTrue(output_embeddings is None or isinstance(output_embeddings, nn.Linear))
+
+    @unittest.skip(reason="VideoMT is not a generative model")
+    def test_generate_without_input_ids(self):
+        pass
+
+    @unittest.skip(reason="VideoMT does not use token embeddings")
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    def test_image_inputs_raise(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        model = VideomtForUniversalSegmentation(config).to(torch_device)
+        model.eval()
+
+        with self.assertRaisesRegex(ValueError, "only supports 5D video inputs"):
+            model(inputs_dict["pixel_values_videos"][:, 0])
+
+    def test_pixel_values_name_raises(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        model = VideomtForUniversalSegmentation(config).to(torch_device)
+        model.eval()
+
+        with self.assertRaisesRegex(ValueError, "Use `pixel_values_videos`"):
+            model(pixel_values=inputs_dict["pixel_values_videos"])
+
+
+@slow
+@require_torch
+@require_vision
+class VideomtForUniversalSegmentationIntegrationTest(unittest.TestCase):
+    instance_model_id = "tue-mps/videomt-dinov2-small-ytvis2019"
+    expected_instance_segments_info = [
+        {"id": 0, "label_id": 13, "score": 0.907032},
+        {"id": 1, "label_id": 7, "score": 0.805882},
+        {"id": 2, "label_id": 13, "score": 0.776713},
+    ]
+    expected_instance_segments_info_frame_1 = [
+        {"id": 0, "label_id": 13, "score": 0.958435},
+        {"id": 1, "label_id": 7, "score": 0.79756},
+        {"id": 2, "label_id": 13, "score": 0.893168},
+    ]
+    expected_panoptic_segments_info = [{"id": 0, "label_id": 13, "score": 0.927756}]
+    expected_panoptic_segments_info_frame_1 = [
+        {"id": 0, "label_id": 13, "score": 0.980277},
+        {"id": 1, "label_id": 13, "score": 0.912077},
+    ]
+
+    def prepare_video(self, num_frames=2):
+        frame = np.array(Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png").convert("RGB"))
+        return [frame.copy() for _ in range(num_frames)]
+
+    def prepare_model_and_inputs(self, model_id, num_frames=2, dtype=None):
+        model_kwargs = {"device_map": "auto"}
+        if dtype is not None:
+            model_kwargs["dtype"] = dtype
+
+        model = VideomtForUniversalSegmentation.from_pretrained(model_id, **model_kwargs)
+        processor = AutoVideoProcessor.from_pretrained(model_id)
+        video_frames = self.prepare_video(num_frames=num_frames)
+        inputs = processor(videos=[video_frames], return_tensors="pt").to(model.device)
+        return model, processor, video_frames, inputs
+
+    def run_inference(self, model_id, num_frames=2, dtype=None):
+        model, processor, video_frames, inputs = self.prepare_model_and_inputs(
+            model_id, num_frames=num_frames, dtype=dtype
+        )
+
+        with torch.inference_mode():
+            outputs = model(**inputs)
+
+        self.assert_common_video_outputs(outputs, model, len(video_frames))
+        return model, processor, video_frames, outputs
+
+    def assert_common_video_outputs(self, outputs, model, num_frames):
+        expected_mask_size = (
+            (model.config.image_size // model.config.patch_size) * (2**model.config.num_upscale_blocks),
+            (model.config.image_size // model.config.patch_size) * (2**model.config.num_upscale_blocks),
+        )
+
+        self.assertEqual(
+            outputs.class_queries_logits.shape, (num_frames, model.config.num_queries, model.config.num_labels + 1)
+        )
+        self.assertEqual(
+            outputs.masks_queries_logits.shape, (num_frames, model.config.num_queries, *expected_mask_size)
+        )
+        self.assertTrue(torch.isfinite(outputs.class_queries_logits.float()).all())
+        self.assertTrue(torch.isfinite(outputs.masks_queries_logits.float()).all())
+
+    def assert_segments_info_close(self, actual_segments_info, expected_segments_info):
+        self.assertEqual(len(actual_segments_info), len(expected_segments_info))
+        for actual, expected in zip(actual_segments_info, expected_segments_info):
+            self.assertEqual(actual["id"], expected["id"])
+            self.assertEqual(actual["label_id"], expected["label_id"])
+            self.assertAlmostEqual(actual["score"], expected["score"], delta=1e-3)
+
+    def test_instance_segmentation_inference(self):
+        _, processor, video_frames, outputs = self.run_inference(self.instance_model_id)
+
+        target_sizes = [frame.shape[:2] for frame in video_frames]
+        results = processor.post_process_instance_segmentation(outputs, target_sizes=target_sizes)
+
+        self.assertEqual(len(results), len(video_frames))
+
+        self.assertEqual(results[0]["segmentation"].shape, video_frames[0].shape[:2])
+        self.assertEqual(results[1]["segmentation"].shape, video_frames[1].shape[:2])
+
+        expected_slice = Expectations(
+            {
+                ("cuda", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 1, 1, 1, 1, 1, 1, 1, 1, -1, -1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    ],
+                    device=results[0]["segmentation"].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 1, 1, 1, 1, 1, 1, 1, 1, -1, -1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    ],
+                    device=results[0]["segmentation"].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(results[0]["segmentation"][24:36, 473:485], expected_slice)
+
+        expected_slice = Expectations(
+            {
+                ("cuda", (8, 6)): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    ],
+                    device=results[1]["segmentation"].device,
+                ),
+                ("cuda", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 1, 1, 1, 1, 1, 1, 1, 1, -1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    ],
+                    device=results[1]["segmentation"].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    ],
+                    device=results[1]["segmentation"].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(results[1]["segmentation"][24:36, 472:484], expected_slice)
+        self.assert_segments_info_close(results[0]["segments_info"], self.expected_instance_segments_info)
+        self.assert_segments_info_close(results[1]["segments_info"], self.expected_instance_segments_info_frame_1)
+
+    def test_semantic_segmentation_inference(self):
+        _, processor, video_frames, outputs = self.run_inference(self.instance_model_id)
+
+        target_sizes = [frame.shape[:2] for frame in video_frames]
+        semantic_results = processor.post_process_semantic_segmentation(outputs, target_sizes=target_sizes)
+
+        self.assertEqual(len(semantic_results), len(video_frames))
+        self.assertEqual(semantic_results[0].shape, video_frames[0].shape[:2])
+        self.assertEqual(semantic_results[1].shape, video_frames[1].shape[:2])
+
+        expected_slice = Expectations(
+            {
+                ("cuda", None): torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                    ],
+                    device=semantic_results[0].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13],
+                    ],
+                    device=semantic_results[0].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(semantic_results[0][1:13, 487:499], expected_slice)
+
+        expected_slice = Expectations(
+            {
+                ("cuda", (8, 6)): torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 0, 0, 0, 0],
+                        [0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=semantic_results[1].device,
+                ),
+                ("cuda", None): torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 0, 0, 0, 0],
+                        [0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 13, 13, 13, 13, 13, 13, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=semantic_results[1].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 0, 0, 0, 0],
+                        [0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 13, 13, 13, 13, 13, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 13, 13, 13, 13, 13, 13, 13, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 13, 13, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 13, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=semantic_results[1].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(semantic_results[1][2:14, 488:500], expected_slice)
+
+    def test_panoptic_segmentation_inference(self):
+        _, processor, video_frames, outputs = self.run_inference(self.instance_model_id)
+
+        target_sizes = [frame.shape[:2] for frame in video_frames]
+        panoptic_results = processor.post_process_panoptic_segmentation(outputs, target_sizes=target_sizes)
+
+        self.assertEqual(len(panoptic_results), len(video_frames))
+        self.assertEqual(panoptic_results[0]["segmentation"].shape, video_frames[0].shape[:2])
+        self.assertEqual(panoptic_results[1]["segmentation"].shape, video_frames[1].shape[:2])
+
+        expected_slice = Expectations(
+            {
+                ("cuda", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=panoptic_results[1]["segmentation"].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=panoptic_results[1]["segmentation"].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(panoptic_results[0]["segmentation"][24:36, 473:485], expected_slice)
+
+        expected_slice = Expectations(
+            {
+                ("cuda", (8, 6)): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=panoptic_results[1]["segmentation"].device,
+                ),
+                ("cuda", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=panoptic_results[1]["segmentation"].device,
+                ),
+                ("xpu", None): torch.tensor(
+                    [
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1],
+                        [-1, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    ],
+                    device=panoptic_results[1]["segmentation"].device,
+                ),
+            }
+        ).get_expectation()
+        torch.testing.assert_close(panoptic_results[1]["segmentation"][24:36, 472:484], expected_slice)
+        self.assert_segments_info_close(panoptic_results[0]["segments_info"], self.expected_panoptic_segments_info)
+        self.assert_segments_info_close(
+            panoptic_results[1]["segments_info"], self.expected_panoptic_segments_info_frame_1
+        )
+
+    @require_torch_gpu
+    def test_instance_segmentation_inference_bf16(self):
+        _, _, _, outputs = self.run_inference(self.instance_model_id, dtype=torch.bfloat16)
+
+        self.assertEqual(outputs.class_queries_logits.dtype, torch.bfloat16)
+        self.assertEqual(outputs.masks_queries_logits.dtype, torch.bfloat16)

tests/models/videomt/test_video_processing_videomt.py

@@ -0,0 +1,191 @@
+# Copyright 2026 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.image_utils import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from transformers.testing_utils import require_torch, require_torchvision, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_video_processing_common import VideoProcessingTestMixin, prepare_video_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    if is_torchvision_available():
+        from transformers import VideomtVideoProcessor
+
+    if is_torch_available():
+        from transformers.models.videomt.modeling_videomt import VideomtForUniversalSegmentationOutput
+
+
+class VideomtVideoProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=5,
+        num_frames=8,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=80,
+        do_resize=True,
+        size=None,
+        do_center_crop=False,
+        do_rescale=True,
+        rescale_factor=1 / 255,
+        do_normalize=True,
+        image_mean=IMAGENET_DEFAULT_MEAN,
+        image_std=IMAGENET_DEFAULT_STD,
+        do_convert_rgb=True,
+        num_queries=3,
+        num_classes=2,
+    ):
+        super().__init__()
+        size = size if size is not None else {"height": 20, "width": 20}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_frames = num_frames
+        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_center_crop = do_center_crop
+        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.do_convert_rgb = do_convert_rgb
+        self.num_queries = num_queries
+        self.num_classes = num_classes
+
+    def prepare_video_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_center_crop": self.do_center_crop,
+            "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,
+            "do_convert_rgb": self.do_convert_rgb,
+        }
+
+    def expected_output_video_shape(self, videos):
+        return self.num_frames, self.num_channels, self.size["height"], self.size["width"]
+
+    def prepare_video_inputs(self, equal_resolution=False, return_tensors="pil"):
+        return prepare_video_inputs(
+            batch_size=self.batch_size,
+            num_frames=self.num_frames,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            return_tensors=return_tensors,
+        )
+
+    def prepare_fake_videomt_outputs(self, num_frames):
+        height, width = self.size["height"], self.size["width"]
+        return VideomtForUniversalSegmentationOutput(
+            masks_queries_logits=torch.randn((num_frames, self.num_queries, height, width)),
+            class_queries_logits=torch.randn((num_frames, self.num_queries, self.num_classes + 1)),
+        )
+
+
+@require_torch
+@require_vision
+@require_torchvision
+class VideomtVideoProcessingTest(VideoProcessingTestMixin, unittest.TestCase):
+    fast_video_processing_class = VideomtVideoProcessor if is_torchvision_available() else None
+    input_name = "pixel_values_videos"
+
+    def setUp(self):
+        super().setUp()
+        self.video_processor_tester = VideomtVideoProcessingTester(self)
+
+    @property
+    def video_processor_dict(self):
+        return self.video_processor_tester.prepare_video_processor_dict()
+
+    def test_video_processor_properties(self):
+        video_processing = self.fast_video_processing_class(**self.video_processor_dict)
+        self.assertTrue(hasattr(video_processing, "do_resize"))
+        self.assertTrue(hasattr(video_processing, "size"))
+        self.assertTrue(hasattr(video_processing, "do_center_crop"))
+        self.assertTrue(hasattr(video_processing, "do_normalize"))
+        self.assertTrue(hasattr(video_processing, "image_mean"))
+        self.assertTrue(hasattr(video_processing, "image_std"))
+        self.assertTrue(hasattr(video_processing, "do_convert_rgb"))
+        self.assertTrue(hasattr(video_processing, "model_input_names"))
+        self.assertIn("pixel_values_videos", video_processing.model_input_names)
+
+    def test_video_processor_from_dict_with_kwargs(self):
+        video_processor = self.fast_video_processing_class.from_dict(self.video_processor_dict)
+        self.assertEqual(video_processor.size, {"height": 20, "width": 20})
+
+        video_processor = self.fast_video_processing_class.from_dict(self.video_processor_dict, size=42)
+        self.assertEqual(video_processor.size, {"height": 42, "width": 42})
+
+    def test_post_process_semantic_segmentation(self):
+        video_processor = self.fast_video_processing_class(**self.video_processor_dict)
+
+        num_frames = 4
+        target_sizes = [(32, 32)] * num_frames
+        outputs = self.video_processor_tester.prepare_fake_videomt_outputs(num_frames)
+
+        segmentation = video_processor.post_process_semantic_segmentation(outputs, target_sizes)
+
+        self.assertEqual(len(segmentation), num_frames)
+        for seg_map in segmentation:
+            self.assertIsInstance(seg_map, torch.Tensor)
+            self.assertEqual(seg_map.shape, (32, 32))
+
+    def test_post_process_instance_segmentation(self):
+        video_processor = self.fast_video_processing_class(**self.video_processor_dict)
+
+        num_frames = 4
+        target_sizes = [(32, 32)] * num_frames
+        outputs = self.video_processor_tester.prepare_fake_videomt_outputs(num_frames)
+
+        results = video_processor.post_process_instance_segmentation(outputs, target_sizes)
+
+        self.assertEqual(len(results), num_frames)
+        for el in results:
+            self.assertIn("segmentation", el)
+            self.assertIn("segments_info", el)
+            self.assertIsInstance(el["segments_info"], list)
+            self.assertEqual(el["segmentation"].shape, (32, 32))
+
+    def test_post_process_panoptic_segmentation(self):
+        video_processor = self.fast_video_processing_class(**self.video_processor_dict)
+
+        num_frames = 4
+        target_sizes = [(32, 32)] * num_frames
+        outputs = self.video_processor_tester.prepare_fake_videomt_outputs(num_frames)
+
+        results = video_processor.post_process_panoptic_segmentation(outputs, target_sizes)
+
+        self.assertEqual(len(results), num_frames)
+        for el in results:
+            self.assertIn("segmentation", el)
+            self.assertIn("segments_info", el)
+            self.assertIsInstance(el["segments_info"], list)
+            self.assertEqual(el["segmentation"].shape, (32, 32))