[OpenVINO] Support Youtu-VL-4B-Instruct with task image-text-to-text

[Mohamed-Ashraf273]

Description

Added OpenVINO export + runtime support for the custom youtu_vl (YoutuVLForConditionalGeneration) image-text-to-text VLM: MLA (DeepSeek/MiniCPM3-style) language model + Qwen2.5-VL-style windowed Siglip2 vision tower split into patch-embed + transformer/merger submodels. Export succeeds (4 submodels) and OV generation matches PyTorch ground truth exactly. Fixed the spatial_shapes generate() blocker. Added and locally verified VLM integration, 4-bit CLI, and auto-compression tests, plus docs entry.

Conversion

optimum-cli export openvino -m /home/mohamed-ashraf/Desktop/projects/GSoC26/auto-openvino-bot/workspace/tiny-youtu-vl ov_youtu --task image-text-to-text --trust-remote-code

Reproduce generation

import torch
from PIL import Image
from transformers import AutoProcessor
from optimum.intel.openvino import OVModelForVisualCausalLM

model_id = '/home/mohamed-ashraf/Desktop/projects/GSoC26/auto-openvino-bot/workspace/tiny-youtu-vl'
ov_dir = 'ov_youtu'

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = OVModelForVisualCausalLM.from_pretrained(ov_dir, trust_remote_code=True)

image = Image.new('RGB', (64, 64), (120, 120, 120))
messages = [{'role': 'user', 'content': [{'type': 'image'}, {'type': 'text', 'text': 'What is this?'}]}]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(images=image, text=prompt, return_tensors='pt')

output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
print(processor.decode(output[0, inputs['input_ids'].shape[1]:], skip_special_tokens=True))

Validation

• WWB similarity: 1.0

Related model-support PRs

• OpenVINO GenAI: [OpenVINO] Support Youtu-VL-4B-Instruct with task image-text-to-text openvinotoolkit/openvino.genai#4046

Before submitting

• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Changed files

• docs/source/openvino/models.mdx
• optimum/exporters/openvino/model_configs.py
• optimum/exporters/openvino/model_patcher.py
• optimum/exporters/openvino/utils.py
• optimum/intel/openvino/modeling_visual_language.py
• tests/openvino/test_exporters_cli.py
• tests/openvino/test_quantization.py
• tests/openvino/test_seq2seq.py
• tests/openvino/utils_tests.py

Related model-support PRs

• OpenVINO GenAI: [OpenVINO] Support Youtu-VL-4B-Instruct with task image-text-to-text openvinotoolkit/openvino.genai#4046

Related model-support PRs

• OpenVINO GenAI: [OpenVINO] Support Youtu-VL-4B-Instruct with task image-text-to-text openvinotoolkit/openvino.genai#4046

[rkazants]

please create tiny model during test. Let us not upload tiny model to HF space if this is new arch.
You can check kokoro model

[Mohamed-Ashraf273]

Done!

[rkazants]

this is trust-remote code model, we should have accurate MIN/Max versions. Let us retrieve it from model config file