minicpm fix for bf16 (#1143)

Author: eaidova

optimum/exporters/openvino/model_configs.py

@@ -95,6 +95,7 @@
     LlavaImageEmbeddingModelPatcher,
     LlavaQwen2ImageEmbeddingsModelPatcher,
     MiniCPM3Patcher,
+    MiniCPMModelPatcher,
     MiniCPMVImageEmbeddingsModelPatcher,
     MiniCPMVResamplerModelPatcher,
     MistralModelPatcher,
@@ -221,6 +222,11 @@ class MiniCPMOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
 
+    def patch_model_for_export(
+        self, model: Union["PreTrainedModel", "TFPreTrainedModel"], model_kwargs: Optional[Dict[str, Any]] = None
+    ) -> "ModelPatcher":
+        return MiniCPMModelPatcher(self, model, model_kwargs=model_kwargs)
+
 
 class OVMiniCPM3DummyPastKeyValuesGenerator(MistralDummyPastKeyValuesGenerator):
     def __init__(

optimum/exporters/openvino/model_patcher.py

@@ -3905,3 +3905,18 @@ def __exit__(self, exc_type, exc_value, traceback):
             for layer in self._model.layers:
                 if hasattr(layer.self_attn, "_orig_forward"):
                     layer.self_attn.forward = layer.self_attn._orig_forward
+
+
+class MiniCPMModelPatcher(DecoderModelPatcher):
+    def __init__(
+        self,
+        config: "OnnxConfig",
+        model: Union["PreTrainedModel", "TFPreTrainedModel"],
+        model_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        for layer in model.model.layers:
+            if hasattr(layer, "scale_depth"):
+                layer.self_attn.o_proj.to(torch.float32)
+                layer.mlp.down_proj.to(torch.float32)
+
+        super().__init__(config, model, model_kwargs)