Fix: set _tied_weights_keys=None to avoid BN buffer/parameter confusion in transformers 5.2

modeling_rtmdet.py

@@ -1758,6 +1758,9 @@ class RTMDetModel(PreTrainedModel):
     config_class = RTMDetConfig
     base_model_prefix = "rtmdet"
     main_input_name = "pixel_values"
+    # share_conv aliases BN modules across strides; disable tied-weights
+    # machinery so transformers doesn't try get_parameter() on BN buffers.
+    _tied_weights_keys = None
     
     def __init__(self, config):
         super().__init__(config)