[Model][3/N] Automatic conversion of CrossEncoding model (#20168)

Signed-off-by: wang.yuqi <noooop@126.com>

vllm/config.py

@@ -466,6 +466,9 @@ class ModelConfig:
                     "affect the random state of the Python process that "
                     "launched vLLM.", self.seed)
 
+        # Keep set served_model_name before maybe_model_redirect(self.model)
+        self.served_model_name = get_served_model_name(self.model,
+                                                       self.served_model_name)
         self.model = maybe_model_redirect(self.model)
         # The tokenizer is consistent with the model by default.
         if self.tokenizer is None:
@@ -609,8 +612,6 @@ class ModelConfig:
 
         self.original_max_model_len = self.max_model_len
         self.max_model_len = self.get_and_verify_max_len(self.max_model_len)
-        self.served_model_name = get_served_model_name(self.model,
-                                                       self.served_model_name)
         self.multimodal_config = self._init_multimodal_config()
         if not self.skip_tokenizer_init:
             self._verify_tokenizer_mode()
@@ -1420,7 +1421,7 @@ class ModelConfig:
 
     @property
     def is_cross_encoder(self) -> bool:
-        return self.registry.is_cross_encoder_model(self.architectures)
+        return self.task == "classify"
 
     @property
     def use_mla(self) -> bool:
@@ -4762,6 +4763,12 @@ class VllmConfig:
         if cls is not None:
             cls.verify_and_update_config(self)
 
+        if self.model_config.task == "classify":
+            # Maybe convert ForCausalLM into ForSequenceClassification model.
+            from vllm.model_executor.models.adapters import (
+                SequenceClassificationConfig)
+            SequenceClassificationConfig.verify_and_update_config(self)
+
     def __str__(self):
         return (
             f"model={self.model_config.model!r},"