fix: hardcode the model max length for vllm

Author: terryyz

bigcodebench/provider/vllm.py

@@ -41,7 +41,8 @@ def __init__(self, name: str, lora_path: str, dataset: str, tp: int, **kwargs) -
                 local_lora_path,
             )
 
-        self.llm = LLM(model=name, enable_lora=True if self.lora_request else False, **kwargs)
+        # max_model_len is set to max_new_tokens * 10
+        self.llm = LLM(model=name, max_model_len=self.max_new_tokens * 10, enable_lora=True if self.lora_request else False, **kwargs)
         self.llm.set_tokenizer(tokenizer=self.tokenizer)
 
     def is_direct_completion(self) -> bool: