change doc

docs/source/use-vllm-as-backend.mdx

@@ -29,6 +29,8 @@ lighteval vllm \
     "leaderboard|truthfulqa:mc|0|0"
 ```
 
+## Use a config file
+
 For more advanced configurations, you can use a config file for the model.
 An example of a config file is shown below and can be found at `examples/model_configs/vllm_model_config.yaml`.
 

examples/model_configs/vllm_model_config.yaml

@@ -7,6 +7,8 @@ model:
     frequency_penalty: 0.0
     presence_penalty: 0.0
     seed: 42
-    top_k: 0
+    top_k: -1
     min_p: 0.0
     top_p: 0.9
+    max_new_tokens: 100
+    stop_tokens: ["<EOS>", "<PAD>"]

src/lighteval/models/model_input.py

@@ -59,6 +59,23 @@ def from_dict(cls, config_dict: dict):
         """
         return GenerationParameters(**config_dict.get("generation", {}))
 
+    def to_vllm_dict(self) -> dict:
+        """Selects relevant generation and sampling parameters for vllm models.
+        Doc: https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html
+
+        Returns:
+            dict: The parameters to create a vllm.SamplingParams in the model config.
+        """
+        sampling_params_to_vllm_naming = {
+            "max_new_tokens": "max_tokens",
+            "min_new_tokens": "min_tokens",
+            "stop_tokens": "stop",
+        }
+
+        # Task specific sampling params to set in model: n, best_of, use_beam_search
+        # Generation specific params to set in model: logprobs, prompt_logprobs
+        return {sampling_params_to_vllm_naming.get(k, k): v for k, v in asdict(self).items() if v is not None}
+
     def to_vllm_openai_dict(self) -> dict:
         """Selects relevant generation and sampling parameters for vllm and openai models.
         Doc: https://docs.vllm.ai/en/v0.5.5/dev/sampling_params.html

src/lighteval/models/vllm/vllm_model.py

@@ -128,7 +128,7 @@ def __init__(
         self.precision = _get_dtype(config.dtype, config=self._config)
 
         self.model_info = ModelInfo(model_name=self.model_name, model_sha=self.model_sha)
-        self.sampling_params = SamplingParams(**config.generation_parameters.to_vllm_openai_dict())
+        self.sampling_params = SamplingParams(**config.generation_parameters.to_vllm_dict())
         self.pairwise_tokenization = config.pairwise_tokenization
 
     @property