Update modeling_qwen.py

modeling_qwen.py

@@ -565,10 +565,13 @@ class QWenModel(QWenPreTrainedModel):
             images = self.visual.encode(images)
             assert images.shape[0] == len(images)
             fake_images = None
-        else:
+        elif self.training:
             fake_images=torch.zeros(1,3,224,224).to(
                 dtype=self.visual.conv1.weight.dtype, device=self.visual.conv1.weight.device)
             images = self.visual(fake_images)
+        else:
+            fake_images = None
+            images = None
 
         output_attentions = (
             output_attentions
@@ -657,7 +660,7 @@ class QWenModel(QWenPreTrainedModel):
         hidden_states = self.drop(hidden_states).clone()
         if fake_images is not None:
             hidden_states = hidden_states + images.mean()*0
-        else:
+        elif images is not None:
             for idx, (i, a, b) in enumerate(img_pos):
                 hidden_states[i][a + 1 : b] = images[idx]
         output_shape = input_shape + (hidden_states.size(-1),)