[OpenVINO] Support Qwen3.5, Qwen3.5-MoE and Qwen3.6

[rkazants]

What does this PR do?

Re-created PR #1634

Fixes 181271, 181280, 182003

Installation instructions:

pip install -U git+https://github.com/rkazants/optimum-intel.git@support_qwen3_5
pip install --pre -U openvino openvino-tokenizers nncf --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
pip install transformers==5.2.0
pip install requests torchvision opencv-python

Exporting cmd-line:

optimum-cli export openvino -m Qwen/Qwen3.5-0.8B Qwen3.5-0.8B

Inference script:

from transformers import AutoProcessor
from transformers.video_utils import load_video
from huggingface_hub import hf_hub_download
from optimum.intel.openvino import OVModelForVisualCausalLM

model_dir = "Qwen/Qwen3.5-0.8B"

processor = AutoProcessor.from_pretrained(model_dir)
model = OVModelForVisualCausalLM.from_pretrained(model_dir)

# Prepare video input
video_path = hf_hub_download(
                repo_id="raushan-testing-hf/videos-test",
                filename="sample_demo_1.mp4",
                repo_type="dataset",
            )
input_video, _ = load_video(video_path, num_frames=10, backend="opencv")

messages = [
    {"role": "user", "content": [
        {"type": "video"},
        {"type": "text", "text": "Why is this video funny?"},
    ]}
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], videos=[input_video], return_tensors="pt")

# Run inference
output_ids = model.generate(**inputs, max_new_tokens=100)
output_text = processor.decode(output_ids[0], skip_special_tokens=True)

print(output_text)

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?

[Copilot]

Pull request overview

This PR adds OpenVINO export + runtime support for Qwen3.5-family visual-language models by introducing new OpenVINO configs/patchers and integrating them into the model loading/export paths.

Changes:

• Add Qwen3.5/Qwen3.5-MoE model types to OpenVINO exporter utilities, stateful patching logic, and model patchers.
• Implement an OpenVINO Visual Causal LM wrapper for Qwen3.5 (including position-id handling and vision embedding utilities).
• Update docs and tests metadata to include Qwen3.5.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
tests/openvino/utils_tests.py	Adds tiny test model id and expected INT8 node counts for qwen3_5.
optimum/intel/openvino/modeling_visual_language.py	Adds Qwen3.5 VLM runtime class + position id handling + model-type mapping.
optimum/intel/openvino/modeling_decoder.py	Ensures full-context attention mask behavior for Qwen3.5 text model types during decoding.
optimum/exporters/openvino/utils.py	Registers Qwen3.5 model types for submodel detection and SSM model handling.
optimum/exporters/openvino/stateful.py	Extends stateful patching to handle VLMs whose text_config.model_type is SSM-based.
optimum/exporters/openvino/model_patcher.py	Adds Qwen3.5(+MoE) patchers and patched forward implementations for export.
optimum/exporters/openvino/model_configs.py	Registers new OpenVINO configs + dummy input generators for Qwen3.5 model types.
optimum/exporters/openvino/convert.py	Adds cleanup logic to undo certain 16-bit traceability patches post-conversion.
docs/source/openvino/models.mdx	Adds Qwen3.5 to the supported architectures list.

---

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

[rkazants]

@regisss, let us make regular tests to be passing. Then I will switch on slow tests.
It is needed to avoid HF exceeding downloading rate error when multiple tests are downloading in parallel.

[regisss]

LGTM

[rkazants]

Suggested change

uv pip install "openvino==2026.1.0" "openvino-tokenizers==2026.1.0.0" "openvino-genai==2026.1.0.0"

let us use nightly build

[rkazants]

need to check manually, because there was a regression in some recent nightly build

[popovaan]

Moe block looks the same as LFM2:
https://github.com/huggingface/transformers/blob/v5.0.0/src/transformers/models/lfm2_moe/modeling_lfm2_moe.py#L167
https://github.com/huggingface/transformers/blob/v5.2.0/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py#L823

Can we reuse lfm2 MoE patching?