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

[mlukasze]

⚠️ AUTOMATICALLY GENERATED BY OMEGA AGENT — REQUIRES HUMAN REVIEW ⚠️
This PR was created by an AI agent as part of automated model enablement.
A human maintainer must review and approve it before it can be considered for merge.
Do NOT merge without human review and sign-off.

What does this PR do?

Adds OpenVINO export and inference support for DINOv3 ConvNeXt (dinov3_convnext, e.g. facebook/dinov3-convnext-small-pretrain-lvd1689m).

DINOv3ConvNextModel (native transformers support since ~4.56) is architecturally a standard ConvNeXt CNN backbone (depthwise conv -> channels-last LayerNorm -> pointwise GELU MLP -> layer-scale -> droppath, 4 stages, global-average-pool + final LayerNorm) with a pixel_values -> last_hidden_state/pooler_output signature -- fully compatible with the existing ViTOpenVINOConfig contract already used by plain convnext. This PR follows the exact same one-line pass-through pattern as ConvNextOpenVINOConfig, registered under the feature-extraction task only (this model has no image-classification/masked-im head/variant, unlike plain convnext).

Validation (real full-size gated model, real COCO image, not synthetic data): PyTorch FP32 vs. OpenVINO IR (OVModelForCustomTasks) cosine similarity = 0.99999994 (last_hidden_state, 50x768) / 1.0 (pooled token); max abs diff 3.53e-05 / 7.27e-06. No missing/unsupported OpenVINO operators (Convolution, GroupConvolution, MVN, MatMul, Add, Multiply, Transpose, Gelu, ReduceMean, Concat, Reshape -- all standard/long-supported).

Test coverage: extended OVModelForCustomTasksIntegrationTest with test_compare_output_vision_feature_extraction, parametrized on dinov3_convnext. OVModelForFeatureExtraction is hard-coded for text inputs (input_ids/attention_mask/token_type_ids) so this test uses the generic OVModelForCustomTasks wrapper instead, with the required explicit task="feature-extraction" kwarg (this class has no default export_feature, unlike task-specific OVModelForXXX classes -- omitting it crashes task inference with TypeError: unsupported operand type(s) for +: 'NoneType' and 'str', a real bug caught and fixed while validating this new test).

Known limitation -- tiny CI fixture not yet uploaded: tests/openvino/utils_tests.py registers optimum-intel-internal-testing/tiny-random-dinov3_convnext per convention, but this repo does not exist on the Hub yet, and the automation account used to prepare this PR only has read-scoped Hub access (confirmed via a 403 on repo creation, including under its own personal namespace) -- it cannot self-publish this fixture. A helper script is included at the bottom of this description for a maintainer with optimum-intel-internal-testing write access to run once (create_tiny_dinov3_convnext.py --push-to-hub); until then, CI will fail on this one parametrized case with a clear "not a valid model identifier" error (not a code defect). The test logic itself has been independently verified to genuinely pass end-to-end (1 passed in a real local pytest run) by temporarily pointing the fixture entry at the real gated model with an authorized token, then reverting -- see the attached test_run.log.

Tiny fixture creation script (for a maintainer with org write access)

python create_tiny_dinov3_convnext.py --push-to-hub
# pushes to optimum-intel-internal-testing/tiny-random-dinov3_convnext
# (script builds a depths=[1,1,1,1], hidden_sizes=[8,16,32,64] tiny random DINOv3ConvNextModel)

Installation instructions

pip install git+https://github.com/mlukasze/optimum-intel.git@enable/facebook-dinov3-convnext-small-pretrain-lvd1689m
pip install --pre -U openvino openvino-tokenizers nncf --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly

Exporting cmd-line

export HF_TOKEN=<your_token_with_accepted_license>
optimum-cli export openvino --model facebook/dinov3-convnext-small-pretrain-lvd1689m --task feature-extraction --trust-remote-code ov_model/

Inference script

import numpy as np
from transformers import AutoImageProcessor
from optimum.intel import OVModelForCustomTasks
from PIL import Image
import requests

processor = AutoImageProcessor.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
ov_model = OVModelForCustomTasks.from_pretrained("ov_model/", device="CPU")

image = Image.open(requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw)
inputs = processor(images=image, return_tensors="pt")
outputs = ov_model(**inputs)

last_hidden_state = outputs.last_hidden_state  # (1, num_patches+1, hidden_size)
pooled_output = last_hidden_state[:, 0, :]      # pooled/CLS-equivalent embedding

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?