[llm_bench] Add Qwen3-Omni support (MoE and dense models)

.github/workflows/linux.yml

@@ -1074,6 +1074,14 @@ jobs:
           python -m pytest -vs ${{ env.SRC_DIR }}/tests/python_tests/samples/test_tools_llm_benchmark.py -m "samples and transformers_lower_v5"
         env:
           SAMPLES_PY_DIR: "${{ env.SRC_DIR }}/tools"
+      - name: Test LLM Benchmark Tools with Qwen3-Omni model
+        if: ${{ fromJSON(needs.smart_ci.outputs.affected_components).llm_bench }}
+        run: |
+          source ${{ env.INSTALL_DIR }}/setupvars.sh
+          python -m pip install transformers==5.1
+          python -m pytest -vs ${{ env.SRC_DIR }}/tests/python_tests/samples/test_tools_llm_benchmark.py -k "qwen3-omni"
+        env:
+          SAMPLES_PY_DIR: "${{ env.SRC_DIR }}/tools"
 
   genai_nodejs_tests:
     name: Node.js bindings tests

tests/python_tests/samples/conftest.py

@@ -25,6 +25,22 @@
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
+
+def get_transformers_version():
+    """Return (major, minor) of the installed transformers, or None if unavailable."""
+    try:
+        version = metadata.version("transformers")
+    except metadata.PackageNotFoundError:
+        return None
+    try:
+        major, minor = version.split(".")[:2]
+        return int(major), int(minor)
+    except ValueError:
+        return None
+
+
+TRANSFORMERS_VERSION = get_transformers_version()
+
 # Dictionary containing model configurations.
 # Each key is a model identifier, and the value is a dictionary with:
 # - "name": the model's name or path
@@ -197,6 +213,10 @@
         "name": "optimum-intel-internal-testing/tiny-random-qwen3-vl",
         "convert_args": ["--trust-remote-code", "--task", "image-text-to-text"],
     },
+    "tiny-random-qwen3-omni": {
+        "name": "optimum-intel-internal-testing/tiny-random-qwen3-omni",
+        "convert_args": ["--trust-remote-code", "--task", "image-text-to-text"],
+    },
 }
 
 TEST_FILES = {

tests/python_tests/samples/test_tools_llm_benchmark.py

@@ -12,7 +12,8 @@
 from utils.constants import get_ov_cache_converted_models_dir
 from utils.kokoro_test_assets import prepare_tiny_kokoro_model_path
 from utils.kokoro_test_assets import prepare_tiny_kokoro_ov_path
-from conftest import SAMPLES_PY_DIR, convert_model, download_test_content
+from conftest import SAMPLES_PY_DIR, convert_model, download_test_content, TRANSFORMERS_VERSION
+
 
 convert_draft_model = convert_model
 download_mask_image = download_test_content
@@ -714,3 +715,31 @@ def test_python_tool_llm_benchmark_video_gen_json(
             generate_llm_bench_input_generation_jsonl,
         ] + sample_args
         run_sample(benchmark_py_command)
+
+    @pytest.mark.samples
+    @pytest.mark.skipif(
+        TRANSFORMERS_VERSION is None or TRANSFORMERS_VERSION < (5, 1),
+        reason="Requires transformers >= 5.1",
+    )
+    @pytest.mark.parametrize("download_test_content", ["cat.png"], indirect=True)
+    @pytest.mark.parametrize("convert_model", ["tiny-random-qwen3-omni"], indirect=True)
+    @pytest.mark.parametrize(
+        "sample_args",
+        [
+            ["-d", "cpu", "-n", "1", "--num_steps", "4", "--task", "visual_text_gen", "--optimum"],
+            ["-d", "cpu", "-n", "1", "--num_steps", "4", "--task", "visual_text_gen", "--genai"],
+        ],
+    )
+    def test_python_tool_llm_benchmark_visual_text_gen(self, download_test_content, convert_model, sample_args):
+        benchmark_script = SAMPLES_PY_DIR / "llm_bench/benchmark.py"
+        benchmark_py_command = [
+            sys.executable,
+            benchmark_script,
+            "-m",
+            convert_model,
+            "--media",
+            download_test_content,
+            "--prompt",
+            "What animal is in this image?",
+        ] + sample_args
+        run_sample(benchmark_py_command)

tools/llm_bench/llm_bench_utils/config_class.py

@@ -171,7 +171,7 @@ class UseCaseTextToSpeech(UseCase):
                 "videochat-flash-qwen",
                 "gemma4",
             ]
-        )
+        ),
     ],
     "speech_to_text": [UseCaseSpeech2Text(["whisper"])],
     "image_cls": [UseCaseImageCls(["vit"])],
@@ -251,5 +251,12 @@ class UseCaseTextToSpeech(UseCase):
     "text_to_speech": [UseCaseTextToSpeech(["speecht5", "kokoro"])],
 }
 
+try:
+    from optimum.intel.openvino import OVModelForMultimodalLM
+except ImportError:
+    pass
+else:
+    USE_CASES["visual_text_gen"].append(UseCaseVLM(["qwen3-omni"], ov_cls=OVModelForMultimodalLM))
+
 PA_ATTENTION_BACKEND = "PA"
 SDPA_ATTENTION_BACKEND = "SDPA"