fix conversion for text embeddings for fp16 models (#968)

* fix conversion for text embeddings for fp16 models

* fix rebasing issue

* apply review comments

* Update tests/openvino/utils_tests.py

optimum/exporters/openvino/model_configs.py

@@ -69,6 +69,7 @@
     GptNeoxJapaneseModelPatcher,
     GptNeoxModelPatcher,
     IBertModelPatcher,
+    InputEmbeddingPatcher,
     InternLM2Patcher,
     InternLMModelPatcher,
     InternVLChatImageEmbeddingModelPatcher,
@@ -1264,6 +1265,12 @@ def rename_ambiguous_inputs(self, inputs):
         model_inputs["input"] = inputs["input_ids"]
         return model_inputs
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        # making 16bit tracable overrides embeedings input signature these changes required to prevent this issue
+        return InputEmbeddingPatcher(self, model, model_kwargs)
+
 
 class LlavaConfigBehavior(str, enum.Enum):
     LANGUAGE = "language"

optimum/exporters/openvino/model_patcher.py

@@ -2991,3 +2991,24 @@ def __init__(
     def __exit__(self, exc_type, exc_value, traceback):
         super().__exit__(exc_type, exc_value, traceback)
         self._model.forward = self._model.__orig_forward
+
+
+class InputEmbeddingPatcher(ModelPatcher):
+    def __init__(
+        self,
+        config: "OnnxConfig",
+        model: Union["PreTrainedModel", "TFPreTrainedModel"],
+        model_kwargs: Dict[str, Any],
+    ):
+        model.__orig_forward = model.forward
+
+        def forward(self, input):
+            return self.__orig_forward(input)
+
+        model.forward = types.MethodType(forward, model)
+
+        super().__init__(config, model, model_kwargs)
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        super().__exit__(exc_type, exc_value, traceback)
+        self._model.forward = self._model.__orig_forward