[Bugfix] support tie_word_embeddings for all models (#5724)

vllm/model_executor/models/minicpmv.py

@@ -496,6 +496,10 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal):
         quant_config: Optional[QuantizationConfig] = None,
     ):
         super().__init__()
+        # All MiniCPM-V models disable `tie_word_embeddings` but
+        # `PretrainedConfig.tie_word_embeddings` defaults to True; we cannot
+        # check `tie_word_embeddings` until vLLM integrate MiniCPM-V model
+        # and config class
         self.config = config
         self.multimodal_config = multimodal_config