Support streaming

modeling_qwen.py

@@ -1026,7 +1026,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
                     seed=-1,
                     **kwargs):
                 outputs.append(token.item())
-                yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore')
+                yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore', keep_image_special=True)
 
         return stream_generator()
 
@@ -1159,4 +1159,4 @@ class RMSNorm(torch.nn.Module):
             return rms_norm(x, self.weight, self.eps)
         else:
             output = self._norm(x.float()).type_as(x)
-            return output * self.weight
+            return output * self.weight