Fix loading of vllm model from files (#533)

* commit

* commit

* Update src/lighteval/main_vllm.py

* commit

* change doc

* change doc

* change doc

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

@@ -29,20 +29,31 @@ lighteval vllm \
     "leaderboard|truthfulqa:mc|0|0"
 ```
 
-Available arguments for `vllm` can be found in the `VLLMModelConfig`:
-
-- **pretrained** (str): HuggingFace Hub model ID name or the path to a pre-trained model to load.
-- **gpu_memory_utilisation** (float): The fraction of GPU memory to use.
-- **revision** (str): The revision of the model.
-- **dtype** (str, None): The data type to use for the model.
-- **tensor_parallel_size** (int): The number of tensor parallel units to use.
-- **data_parallel_size** (int): The number of data parallel units to use.
-- **max_model_length** (int): The maximum length of the model.
-- **swap_space** (int): The CPU swap space size (GiB) per GPU.
-- **seed** (int): The seed to use for the model.
-- **trust_remote_code** (bool): Whether to trust remote code during model loading.
-- **add_special_tokens** (bool): Whether to add special tokens to the input sequences.
-- **multichoice_continuations_start_space** (bool): Whether to add a space at the start of each continuation in multichoice generation.
+## 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`.
+
+```bash
+lighteval vllm \
+    "examples/model_configs/vllm_model_config.yaml" \
+    "leaderboard|truthfulqa:mc|0|0"
+```
+
+```yaml
+model: # Model specific parameters
+  base_params:
+    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main,dtype=bfloat16" # Model args that you would pass in the command line
+  generation: # Generation specific parameters
+    temperature: 0.3
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0
+    presence_penalty: 0.0
+    seed: 42
+    top_k: 0
+    min_p: 0.0
+    top_p: 0.9
+```
 
 > [!WARNING]
 > In the case of OOM issues, you might need to reduce the context size of the

examples/model_configs/vllm_model_config.yaml

@@ -0,0 +1,14 @@
+model:
+  base_params:
+    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main,dtype=bfloat16" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
+  generation:
+    temperature: 0.3
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0
+    presence_penalty: 0.0
+    seed: 42
+    top_k: -1
+    min_p: 0.0
+    top_p: 0.9
+    max_new_tokens: 100
+    stop_tokens: ["<EOS>", "<PAD>"]

src/lighteval/main_vllm.py

@@ -133,12 +133,13 @@ def vllm(
     if model_args.endswith(".yaml"):
         with open(model_args, "r") as f:
             config = yaml.safe_load(f)["model"]
+        model_args = config["base_params"]["model_args"]
         generation_parameters = GenerationParameters.from_dict(config)
-        model_config = VLLMModelConfig(config, generation_parameters=generation_parameters)
-
     else:
-        model_args_dict: dict = {k.split("=")[0]: k.split("=")[1] if "=" in k else True for k in model_args.split(",")}
-        model_config = VLLMModelConfig(**model_args_dict)
+        generation_parameters = GenerationParameters()
+
+    model_args_dict: dict = {k.split("=")[0]: k.split("=")[1] if "=" in k else True for k in model_args.split(",")}
+    model_config = VLLMModelConfig(**model_args_dict, generation_parameters=generation_parameters)
 
     pipeline = Pipeline(
         tasks=tasks,

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

@@ -129,7 +129,7 @@ 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