properly fix quantization

Author: eaidova

optimum/intel/openvino/modeling_visual_language.py

@@ -1738,6 +1738,8 @@ def get_multimodal_embeddings(
         vision_embeds = None
         IGNORE_INDEX = -100
         IMAGE_TOKEN_INDEX = -200
+        if pixel_values is None and "images" in kwargs:
+            pixel_values = kwargs["images"]
         if pixel_values is not None:
             vision_embeds = self.get_vision_embeddings(pixel_values, input_ids=input_ids, **kwargs)
         if vision_embeds is None:

tests/openvino/test_exporters_cli.py

@@ -197,7 +197,7 @@ class OVCLIExportTestCase(unittest.TestCase):
                     "image-text-to-text",
                     "nanollava",
                     "int4 --group-size 8 --ratio 0.8 --trust-remote-code",
-                    [{"int8": 14, "int4": 16}, {"int8": 15}, {"int8": 1}],
+                    [{"int8": 16, "int4": 14}, {"int8": 15}, {"int8": 1}],
                 ),
                 (
                     "image-text-to-text",

tests/openvino/test_quantization.py

@@ -466,7 +466,7 @@ class OVWeightCompressionTest(unittest.TestCase):
                         tokenizer=MODEL_NAMES["nanollava"],
                         trust_remote_code=True,
                     ),
-                    [{"int8": 14, "int4": 16}, {"int8": 15}, {"int8": 1}],
+                    [{"int8": 16, "int4": 14}, {"int8": 15}, {"int8": 1}],
                 ),
             ]
         )