[OpenVINO] Support Gemma 4

[rkazants]

What does this PR do?

Fixes 182357

Re-created from PR #1675

Installation instructions:

pip install git+https://github.com/rkazants/optimum-intel.git@support_gemma_4
pip install --pre -U openvino openvino-tokenizers nncf --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
pip install transformers==5.5.0
pip install requests torchvision requests

Exporting cmd-line:

optimum-cli export openvino -m google/gemma-4-E2B-it ov_gemma4_E2Bit --task=image-text-to-text

Inference script:

from transformers import AutoProcessor
import torch
from transformers import AutoProcessor, Gemma4ForConditionalGeneration
from optimum.intel.openvino import OVModelForVisualCausalLM

model_id = "google/gemma-4-E2B-it"
model = OVModelForVisualCausalLM.from_pretrained(model_id)

processor = AutoProcessor.from_pretrained(model_id, padding_side="left")

url = "https://media.istockphoto.com/id/1192867753/photo/cow-in-berchida-beach-siniscola.jpg?s=612x612&w=0&k=20&c=v0hjjniwsMNfJSuKWZuIn8pssmD5h5bSN1peBd1CmH4="
messages = [
    {
        "role": "system",
        "content": [
            {"type": "text", "text": "You are a helpful assistant."}
        ]
    },
    {
        "role": "user", "content": [
            {"type": "image", "url": url},
            {"type": "text", "text": "What is shown in this image?"},
        ]
    },
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
)

output = model.generate(**inputs, max_new_tokens=50)
print(processor.decode(output[0, inputs.input_ids.shape[1]: ], skip_special_tokens=True))

Before submitting

• [N/A] 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?

[echarlaix]

Thanks for the addition!

[echarlaix]

when will this be the case ?

[popovaan]

It's the additional input of the language model for Gemma4. In original Gemma4 they are calculated inside of language model forward, but to calculate them we need input_ids, which we don't pass to language model, so we decided to move them to the separate model for the minimal API impact.

[regisss]

Adding a few comments to @echarlaix's