fix cross_encoding

Author: noooop

vllm/model_executor/models/roberta.py

@@ -178,6 +178,12 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
 
+        if vllm_config.model_config.task == "score":
+            # fix cross_encoding. See #19201
+            padding_idx = getattr(config,
+                                  "pad_token_id", 0)
+            vllm_config.model_config.max_model_len -= padding_idx * 2
+
         self.default_activation_function = \
             get_cross_encoder_activation_function(config)