fix: hardcode the model max length for vllm

terryyz · terryyz · commit 3513d997f55c · 2025-03-25T00:13:39.000+08:00
diff --git a/bigcodebench/provider/vllm.py b/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:

Original file line number	Diff line number	Diff line change
`@@ -41,7 +41,8 @@ def __init__(self, name: str, lora_path: str, dataset: str, tp: int, **kwargs) -`
`41`	`41`	`local_lora_path,`
`42`	`42`	`)`
`43`	`43`
`44`		`- self.llm = LLM(model=name, enable_lora=True if self.lora_request else False, **kwargs)`
	`44`	`+ # max_model_len is set to max_new_tokens * 10`
	`45`	`+ self.llm = LLM(model=name, max_model_len=self.max_new_tokens * 10, enable_lora=True if self.lora_request else False, **kwargs)`
`45`	`46`	`self.llm.set_tokenizer(tokenizer=self.tokenizer)`
`46`	`47`
`47`	`48`	`def is_direct_completion(self) -> bool:`