for ms-swift

modeling_mplugowl3.py

@@ -88,9 +88,8 @@ class mPLUGOwl3Model(mPLUGOwl3PreTrainedModel):
         if pixel_values is None:
             return None
         dtype = self.language_model.model.embed_tokens.weight.dtype
-        with torch.inference_mode():
-            image_embeds = self._small_batched_forward(pixel_values.to(dtype))
-            # image_embeds = self.vision_model(pixel_values.to(dtype), output_hidden_states=True).hidden_states[-2]
+        image_embeds = self._small_batched_forward(pixel_values.to(dtype))
+        # image_embeds = self.vision_model(pixel_values.to(dtype), output_hidden_states=True).hidden_states[-2]
         
         if self.vision2text_model is not None:
             image_embeds = self.vision2text_model(image_embeds)