Update modeling_qwen3_pseudo_moe.py

Fix loading under Transformers 5 (_tied_weights_keys)
Loading fails on Transformers 5.x in post_init() with AttributeError: 'list' object has no attribute 'keys'. The cause is that _tied_weights_keys = [] uses the old list format while config.json sets tie_word_embeddings: true. This sets it to None, the PreTrainedModel default; nothing is tied, since the model has no LM head. It also replaces the deprecated config.use_return_dict with an equivalent expression. Tested with Transformers 4.51.3, 4.57.1 and 5.18.0: all 475 weights load and the embeddings are identical to the original code's.

modeling_qwen3_pseudo_moe.py

@@ -92,7 +92,9 @@ class MultiLoRALinear(nn.Module):
 class Qwen3PseudoMoEModelModel(Qwen3PreTrainedModel):
     config_class = Qwen3PseudoMoEConfig
     _no_split_modules = ["Qwen3DecoderLayer"]
-    _tied_weights_keys = []
+    # Nothing is tied (the model has no LM head). None is the PreTrainedModel default in every Transformers
+    # version; the former [] crashes Transformers 5, which expects a dict here when tie_word_embeddings is true.
+    _tied_weights_keys = None
 
     def __init__(self, config: Qwen3PseudoMoEConfig):
         super().__init__(config)
@@ -169,7 +171,10 @@ class Qwen3PseudoMoEModelModel(Qwen3PreTrainedModel):
         if domain is not None:
             self.set_domain(domain)
 
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        # Same result as config.use_return_dict, which Transformers 5 deprecates (5.x also drops torchscript).
+        return_dict = return_dict if return_dict is not None else (
+            getattr(self.config, "return_dict", True) and not getattr(self.config, "torchscript", False)
+        )
 
         outputs = self.model(
             input_ids=input_ids,
@@ -222,4 +227,4 @@ class Qwen3PseudoMoEModelModel(Qwen3PreTrainedModel):
 
 
 AutoConfig.register("qwen3_pseudo_moe", Qwen3PseudoMoEConfig)
-AutoModel.register(Qwen3PseudoMoEConfig, Qwen3PseudoMoEModelModel)
+AutoModel.register(Qwen3PseudoMoEConfig, Qwen3PseudoMoEModelModel)