Fix: AutoModelForImageTextToText, arXiv preprint, _tied_weights_keys dict

modeling_nas_child_vl.py

@@ -28,7 +28,7 @@ class Qwen3VLCausalLMOutputWithPast(CausalLMOutputWithPast):
 class NasChildVLModelForCausalLM(Qwen3VLPreTrainedModel, GenerationMixin):
     config_class = NasChildVLConfig
     _checkpoint_conversion_mapping = {}
-    _tied_weights_keys = ["lm_head.weight"]
+    _tied_weights_keys = {"lm_head.weight": "embed_tokens.weight"}
 
     get_image_features   = Qwen3VLModel.get_image_features
     get_video_features   = Qwen3VLModel.get_video_features