Fixes to Loglikelihood prefix token / VLLM (#1611)

* make vllm use prefix_token_id ; have prefix_token_id be optional method to define

* custom_prefix_token_id wasn't set if not passed

lm_eval/api/model.py

@@ -284,10 +284,9 @@ class TemplateLM(LM):
         pass
 
     @property
-    @abc.abstractmethod
     def prefix_token_id(self):
         # it is used as prefix for loglikelihood
-        pass
+        return self.eot_token_id
 
     @abc.abstractmethod
     def tok_encode(self, string: str, **kwargs):

lm_eval/models/huggingface.py

@@ -99,6 +99,7 @@ class HFLM(TemplateLM):
         trust_remote_code: Optional[bool] = False,
         use_fast_tokenizer: Optional[bool] = True,
         add_bos_token: Optional[bool] = False,
+        prefix_token_id: Optional[int] = None,
         # arguments used for splitting a model across GPUs naively.
         # only used if `parallelize=True`.
         parallelize: Optional[bool] = False,
@@ -109,7 +110,6 @@ class HFLM(TemplateLM):
         # PEFT and quantization options
         peft: Optional[str] = None,
         autogptq: Optional[Union[bool, str]] = False,
-        prefix_token_id: Optional[int] = None,
         **kwargs,
     ) -> None:
         super().__init__()
@@ -342,6 +342,7 @@ class HFLM(TemplateLM):
             self._world_size = 1
 
         self.custom_prefix_token_id = prefix_token_id
+        if prefix_token_id is not None:
             eval_logger.info(
                 f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
             )

lm_eval/models/vllm_causallms.py

@@ -42,6 +42,7 @@ class VLLM(TemplateLM):
         tokenizer_mode: Literal["auto", "slow"] = "auto",
         tokenizer_revision: Optional[str] = None,
         add_bos_token: Optional[bool] = False,
+        prefix_token_id: Optional[int] = None,
         tensor_parallel_size: int = 1,
         quantization: Optional[str] = None,
         max_gen_toks: int = 256,
@@ -118,6 +119,11 @@ class VLLM(TemplateLM):
             tokenizer_revision=tokenizer_revision,
         )
         self.add_bos_token = add_bos_token
+        self.custom_prefix_token_id = prefix_token_id
+        if prefix_token_id is not None:
+            eval_logger.info(
+                f"Loglikelihood prefix token id used in evaluation: {self.prefix_token_id}"
+            )
 
         self._max_gen_toks = max_gen_toks
 
@@ -126,6 +132,15 @@ class VLLM(TemplateLM):
         # we use EOT because end of *text* is more accurate for what we're doing than end of *sentence*
         return self.tokenizer.eos_token_id
 
+    @property
+    def prefix_token_id(self):
+        # it is used as prefix for loglikelihood
+        if self.custom_prefix_token_id is not None:
+            return self.custom_prefix_token_id
+        if self.tokenizer.bos_token_id is not None:
+            return self.tokenizer.bos_token_id
+        return self.tokenizer.eos_token_id
+
     @property
     def max_length(self):
         if self._max_length:  # if max length manually set, return it

tests/models/test_vllm.py

@@ -25,8 +25,8 @@ class TEST_VLLM:
     multiple_choice_task.build_all_requests(limit=10, rank=0, world_size=1)
     MULTIPLE_CH: List[Instance] = multiple_choice_task.instances
     generate_until_task = task_list["gsm8k"]  # type: ignore
-    generate_until_task.build_all_requests(limit=10, rank=0, world_size=1)
     generate_until_task._config.generation_kwargs["max_gen_toks"] = 10
+    generate_until_task.build_all_requests(limit=10, rank=0, world_size=1)
     generate_until: List[Instance] = generate_until_task.instances
     rolling_task = task_list["wikitext"]  # type: ignore
     rolling_task.build_all_requests(limit=10, rank=0, world_size=1)