fix config

myolmoe/modeling_myolmoe.py

@@ -79,13 +79,24 @@ class MyOlmoeConfig(PretrainedConfig):
     # Model architecture
     architectures: List[str] = field(default_factory=lambda: ["MyOlmoeForCausalLM"])
     
+    def __init__(self, **kwargs):
+        # Remove torch_dtype and other model loading parameters that shouldn't be in config
+        model_loading_params = ['torch_dtype', 'device_map', 'low_cpu_mem_usage']
+        for param in model_loading_params:
+            kwargs.pop(param, None)
+        
+        # Initialize dataclass fields
+        for field in self.__dataclass_fields__:
+            if field in kwargs:
+                setattr(self, field, kwargs.pop(field))
+        
+        # Call parent init with remaining kwargs
+        super().__init__(**kwargs)
+    
     def __post_init__(self):
         """Post-initialization to ensure compatibility with PretrainedConfig."""
-        super().__init__(
-            pad_token_id=self.pad_token_id,
-            eos_token_id=self.eos_token_id,
-            **{k: v for k, v in self.__dict__.items() if not k.startswith('_')}
-        )
+        # This is handled in __init__ now
+        pass
 
 logger = logging.get_logger(__name__)