diff --git a/vllm/config/model.py b/vllm/config/model.py
index 92cd48402a65d..b595c8550491b 100644
--- a/vllm/config/model.py
+++ b/vllm/config/model.py
@@ -1726,19 +1726,12 @@ class ModelConfig:
         logger.debug_once("head dtype: %s", head_dtype)
         return head_dtype
 
-    @property
-    def hidden_size(self):
-        if hasattr(self.hf_config, "hidden_size"):
-            return self.hf_config.hidden_size
-        text_config = self.hf_config.get_text_config()
-        return text_config.hidden_size
-
     @property
     def embedding_size(self):
         dense_modules = try_get_dense_modules(self.model, revision=self.revision)
         if dense_modules is not None:
             return dense_modules[-1]["out_features"]
-        return self.hidden_size
+        return self.get_hidden_size()
 
     def get_and_verify_max_len(self, max_model_len: int):
         # Consider max_model_len in tokenizer_config only when
diff --git a/vllm/model_executor/models/adapters.py b/vllm/model_executor/models/adapters.py
index 738400ae864a5..05f257feea3ee 100644
--- a/vllm/model_executor/models/adapters.py
+++ b/vllm/model_executor/models/adapters.py
@@ -301,7 +301,7 @@ def as_seq_cls_model(cls: _T) -> _T:
             quant_config = vllm_config.quant_config
 
             self.score = ReplicatedLinear(
-                model_config.hidden_size,
+                model_config.get_hidden_size(),
                 text_config.num_labels,
                 bias=False,
                 params_dtype=vllm_config.model_config.head_dtype,