add test case

optimum/intel/openvino/modeling_visual_language.py

@@ -165,6 +165,10 @@ def prepare_inputs(
             if past_len:
                 position_ids = position_ids[:, -inputs_embeds.shape[1] :]
 
+            if self.config.model_type == "qwen2_vl" and position_ids.ndim != 3:
+                position_ids = np.expand_dims(position_ids, 0)
+                position_ids = np.concatenate([position_ids, position_ids, position_ids], axis=0)
+
             inputs["position_ids"] = position_ids
 
         if "beam_idx" in self.input_names:

tests/openvino/test_modeling.py

@@ -1976,7 +1976,7 @@ class OVModelForVisualCausalLMIntegrationTest(unittest.TestCase):
     if is_transformers_version(">=", "4.40.0"):
         SUPPORTED_ARCHITECTURES += ["llava_next", "nanollava"]
     if is_transformers_version(">=", "4.45.0"):
-        SUPPORTED_ARCHITECTURES += ["minicpmv", "internvl2", "phi3_v"]
+        SUPPORTED_ARCHITECTURES += ["minicpmv", "internvl2", "phi3_v", "qwen2_vl"]
     TASK = "image-text-to-text"
     REMOTE_CODE_MODELS = ["internvl2", "minicpmv", "nanollava", "phi3_v"]
 
@@ -1996,6 +1996,10 @@ def get_transformer_model_class(self, model_arch):
             from transformers import LlavaNextForConditionalGeneration
 
             return LlavaNextForConditionalGeneration
+        if model_arch == "qwen2_vl":
+            from transformers import Qwen2VLForConditionalGeneration
+
+            return Qwen2VLForConditionalGeneration
         return AutoModelForCausalLM
 
     def _check_device_and_request(self, ov_model, expected_device, has_request):

tests/openvino/utils_tests.py

@@ -115,6 +115,7 @@
     "qwen": "katuni4ka/tiny-random-qwen",
     "qwen2": "fxmarty/tiny-dummy-qwen2",
     "qwen2-moe": "katuni4ka/tiny-random-qwen1.5-moe",
+    "qwen2_vl": "katuni4ka/tiny-random-qwen2vl",
     "resnet": "hf-internal-testing/tiny-random-resnet",
     "roberta": "hf-internal-testing/tiny-random-roberta",
     "roformer": "hf-internal-testing/tiny-random-roformer",