Add support for Ouro (ByteDance/Ouro-1.4B)

[openvino-dev-samples]

What does this PR do?

This PR adds OpenVINO export and inference support for the Ouro model family (OuroForCausalLM).

Ouro is a Universal Transformer decoder: the same num_hidden_layers decoder layers are looped total_ut_steps times, and every iteration stores its own key/value entry. The only thing required for a correct export is a normalized config that reports num_layers = num_hidden_layers * total_ut_steps, so the exported model exposes the right number of past-key-value pairs. No custom model patching or new operators are needed — the stock OVDecoderModelPatcher is reused.

Export the model:

optimum-cli export openvino --model ByteDance/Ouro-1.4B --trust-remote-code ouro_ov

Run inference:

from transformers import AutoTokenizer
from optimum.intel import OVModelForCausalLM

model_id = "ByteDance/Ouro-1.4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = OVModelForCausalLM.from_pretrained(model_id, export=True, trust_remote_code=True)

inputs = tokenizer("The quick brown fox", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

Notes:

• Ouro relies on remote modeling code that is incompatible with transformers v5, so tests are gated to >=4.53.0, <5 (validated against transformers==4.57.1).
• Ouro's custom UniversalTransformerCache makes cached generation diverge from uncached under left-padding + beam search (this happens in transformers too), so in tests Ouro joins the other remote-code custom-cache models that compare against transformers with use_cache=False.
• A tiny test model is added at optimum-intel-internal-testing/tiny-random-ouro (full vocab + real tokenizer, total_ut_steps=4).

Fixes # (issue)

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?

[Copilot]

Pull request overview

Adds OpenVINO export/inference support for the ByteDance Ouro Universal Transformer decoder family by introducing a normalized config that reports an effective layer count (to expose the correct number of past key/value pairs), plus wiring the architecture into the OpenVINO test matrix and documentation.

Changes:

• Register ouro in the OpenVINO TasksManager with a custom NormalizedOuroConfig that expands num_layers = num_hidden_layers * total_ut_steps.
• Add ouro to OpenVINO decoder/export/CLI/quantization test parametrizations (gated to transformers>=4.53.0,<5) and tiny model mappings.
• Document Ouro as a supported OpenVINO architecture.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/openvino/utils_tests.py	Adds tiny Ouro test model mapping, expected node counts, and remote-code gating for <5.
tests/openvino/test_quantization.py	Adds Ouro to auto-compression test coverage under the <5 transformers gate.
tests/openvino/test_exporters_cli.py	Adds Ouro to CLI export test matrix and expected tokenizer model count.
tests/openvino/test_export.py	Adds Ouro to exporter integration tests under the <5 transformers gate.
tests/openvino/test_decoder.py	Adds Ouro to decoder integration tests and aligns comparison behavior with other custom-cache remote-code models.
optimum/exporters/openvino/model_configs.py	Registers ouro OpenVINO config and introduces NormalizedOuroConfig to expand effective layer count.
docs/source/openvino/models.mdx	Lists Ouro among supported OpenVINO architectures.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[rkazants]

document this model in the bottom and share links. Check other trust-remote-code models like Kokoro.
Do we need any additinal library to setup for model loading?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[openvino-agent]

since this is trust-remote-code model, what is max version?