allows better flexibility for litellm endpoints (#549)

* allows better flexibility for litellm

* add config file

* add doc

* add doc

* add doc

* add doc

* add lighteval imgs

* add lighteval imgs

* add doc

* Update docs/source/use-litellm-as-backend.mdx

Co-authored-by: Victor Muštar <[email protected]>

* Update docs/source/use-litellm-as-backend.mdx

---------

Co-authored-by: Victor Muštar <[email protected]>

Author: NathanHB

docs/source/_toctree.yml

@@ -9,6 +9,8 @@
 - sections:
   - local: saving-and-reading-results
     title: Save and read results
+  - local: use-litellm-as-backend
+    title: Use LITELLM as backend
   - local: using-the-python-api
     title: Use the Python API
   - local: adding-a-custom-task

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

@@ -0,0 +1,81 @@
+# Litellm as backend
+
+Lighteval allows to use litellm, a backend allowing you to call all LLM APIs
+using the OpenAI format [Bedrock, Huggingface, VertexAI, TogetherAI, Azure,
+OpenAI, Groq etc.].
+
+Documentation for available APIs and compatible endpoints can be found [here](https://docs.litellm.ai/docs/).
+
+## Quick use
+
+```bash
+lighteval endpoint litellm \
+    "gpt-3.5-turbo" \
+    "lighteval|gsm8k|0|0"
+```
+
+## Using a config file
+
+Litellm allows generation with any OpenAI compatible endpoint, for example you
+can evaluate a model running on a local vllm server.
+
+To do so you will need to use a config file like so:
+
+```yaml
+model:
+  base_params:
+    model_name: "openai/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
+    base_url: "URL OF THE ENDPOINT YOU WANT TO USE"
+    api_key: "" # remove or keep empty as needed
+  generation:
+    temperature: 0.5
+    max_new_tokens: 256
+    stop_tokens: [""]
+    top_p: 0.9
+    seed: 0
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0
+```
+
+## Use Hugging Face Inference Providers
+
+With this you can also access HuggingFace Inference servers, let's look at how to evaluate DeepSeek-R1-Distill-Qwen-32B.
+
+First, let's look at how to acess the model, we can find this from [the model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B).
+
+Step 1:
+
+![Step 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lighteval/litellm-guide-2.png)
+
+Step 2:
+
+![Step 2](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lighteval/litellm-guide-1.png)
+
+Great ! Now we can simply copy paste the base_url and our api key to eval our model.
+
+> [!WARNING]
+> Do not forget to prepend the provider in the `model_name`. Here we use an
+> openai compatible endpoint to the provider is `openai`.
+
+```yaml
+model:
+  base_params:
+    model_name: "openai/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
+    base_url: "https://router.huggingface.co/hf-inference/v1"
+    api_key: "YOUR KEY" # remove or keep empty as needed
+  generation:
+    temperature: 0.5
+    max_new_tokens: 256 # This will overide the default from the tasks config
+    top_p: 0.9
+    seed: 0
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0
+```
+
+And then, we are able to eval our model on any eval available in Lighteval.
+
+```bash
+lighteval endpoint litellm \
+    "examples/model_configs/litellm_model.yaml" \
+    "lighteval|gsm8k|0|0"
+```

examples/model_configs/litellm_model.yaml

@@ -0,0 +1,12 @@
+model:
+  base_params:
+    model_name: "openai/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
+    base_url: "https://router.huggingface.co/hf-inference/v1"
+  generation:
+    temperature: 0.5
+    max_new_tokens: 256
+    stop_tokens: [""]
+    top_p: 0.9
+    seed: 0
+    repetition_penalty: 1.0
+    frequency_penalty: 0.0

src/lighteval/main_endpoint.py

@@ -385,8 +385,11 @@ def tgi(
 @app.command(rich_help_panel="Evaluation Backends")
 def litellm(
     # === general ===
-    model_name: Annotated[
-        str, Argument(help="The model name to evaluate (has to be available through the litellm API.")
+    model_args: Annotated[
+        str,
+        Argument(
+            help="config file path for the litellm model, or a comma separated string of model args (model_name={},base_url={},provider={})"
+        ),
     ],
     tasks: Annotated[str, Argument(help="Comma-separated list of tasks to evaluate on.")],
     # === Common parameters ===
@@ -462,7 +465,11 @@ def litellm(
     # TODO (nathan): better handling of model_args
     parallelism_manager = ParallelismManager.NONE
 
-    model_config = LiteLLMModelConfig(model=model_name)
+    if model_args.endswith(".yaml"):
+        model_config = LiteLLMModelConfig.from_path(model_args)
+    else:
+        model_name = model_args.split(",")[0].strip()
+        model_config = LiteLLMModelConfig(model=model_name)
 
     pipeline_params = PipelineParameters(
         launcher_type=parallelism_manager,

src/lighteval/models/litellm_model.py

@@ -21,17 +21,18 @@
 # SOFTWARE.
 
 import logging
-import os
 import time
 from concurrent.futures import ThreadPoolExecutor
 from dataclasses import dataclass
 from typing import Optional
 
+import yaml
 from tqdm import tqdm
 
 from lighteval.data import GenerativeTaskDataset
 from lighteval.models.abstract_model import LightevalModel
 from lighteval.models.endpoints.endpoint_model import ModelInfo
+from lighteval.models.model_input import GenerationParameters
 from lighteval.models.model_output import (
     GenerativeResponse,
     LoglikelihoodResponse,
@@ -63,6 +64,32 @@
 @dataclass
 class LiteLLMModelConfig:
     model: str
+    provider: Optional[str] = None
+    base_url: Optional[str] = None
+    api_key: Optional[str] = None
+    generation_parameters: GenerationParameters = None
+
+    def __post_init__(self):
+        if self.generation_parameters is None:
+            self.generation_parameters = GenerationParameters()
+
+    @classmethod
+    def from_path(cls, path):
+        with open(path, "r") as f:
+            config = yaml.safe_load(f)["model"]
+
+        model = config["base_params"]["model_name"]
+        provider = config["base_params"].get("provider", None)
+        base_url = config["base_params"].get("base_url", None)
+        api_key = config["base_params"].get("api_key", None)
+        generation_parameters = GenerationParameters.from_dict(config)
+        return cls(
+            model=model,
+            provider=provider,
+            base_url=base_url,
+            generation_parameters=generation_parameters,
+            api_key=api_key,
+        )
 
 
 class LiteLLMClient(LightevalModel):
@@ -79,15 +106,17 @@ def __init__(self, config, env_config) -> None:
             model_dtype=None,
             model_size="",
         )
-        self.provider = config.model.split("/")[0]
-        self.base_url = os.getenv(f"{self.provider.upper()}_BASE_URL", None)
+        self.model = config.model
+        self.provider = config.provider or config.model.split("/")[0]
+        self.base_url = config.base_url
+        self.api_key = config.api_key
+        self.generation_parameters = config.generation_parameters
+
         self.API_MAX_RETRY = 5
         self.API_RETRY_SLEEP = 3
         self.API_RETRY_MULTIPLIER = 2
         self.CONCURENT_CALLS = 20  # 100 leads to hitting Anthropic rate limits
-        self.TEMPERATURE = 0.3
-        self.TOP_P = 0.95
-        self.model = config.model
+
         self._tokenizer = encode
         self.pairwise_tokenization = False
         litellm.drop_params = True
@@ -126,18 +155,19 @@ def __call_api(self, prompt, return_logits, max_new_tokens, num_samples, stop_se
                 kwargs = {
                     "model": self.model,
                     "messages": prompt,
-                    "max_completion_tokens": max_new_tokens,
                     "logprobs": return_logits if self.provider == "openai" else None,
                     "base_url": self.base_url,
                     "n": num_samples,
                     "caching": True,
+                    "api_key": self.api_key,
                 }
                 if "o1" in self.model:
                     logger.warning("O1 models do not support temperature, top_p, stop sequence. Disabling.")
                 else:
-                    kwargs["temperature"] = self.TEMPERATURE
-                    kwargs["top_p"] = self.TOP_P
-                    kwargs["stop"] = stop_sequence
+                    kwargs.update(self.generation_parameters.to_litellm_dict())
+
+                if kwargs.get("max_completion_tokens", None) is None:
+                    kwargs["max_completion_tokens"] = max_new_tokens
 
                 response = litellm.completion(**kwargs)
 

src/lighteval/models/model_input.py

@@ -32,15 +32,15 @@ class GenerationParameters:
     length_penalty: Optional[float] = None  # vllm, transformers
     presence_penalty: Optional[float] = None  # vllm
 
-    max_new_tokens: Optional[int] = None  # vllm, transformers, tgi
+    max_new_tokens: Optional[int] = None  # vllm, transformers, tgi, litellm
     min_new_tokens: Optional[int] = None  # vllm, transformers
 
-    seed: Optional[int] = None  # vllm, tgi
-    stop_tokens: Optional[list[str]] = None  # vllm, transformers, tgi
-    temperature: Optional[float] = None  # vllm, transformers, tgi
+    seed: Optional[int] = None  # vllm, tgi litellm
+    stop_tokens: Optional[list[str]] = None  # vllm, transformers, tgi, litellm
+    temperature: Optional[float] = None  # vllm, transformers, tgi, litellm
     top_k: Optional[int] = None  # vllm, transformers, tgi
     min_p: Optional[float] = None  # vllm, transformers
-    top_p: Optional[int] = None  # vllm, transformers, tgi
+    top_p: Optional[int] = None  # vllm, transformers, tgi, litellm
     truncate_prompt: Optional[bool] = None  # vllm, tgi
 
     @classmethod
@@ -59,6 +59,24 @@ def from_dict(cls, config_dict: dict):
         """
         return GenerationParameters(**config_dict.get("generation", {}))
 
+    def to_litellm_dict(self) -> dict:
+        """Selects relevant generation and sampling parameters for litellm models.
+        Doc: https://docs.litellm.ai/docs/completion/input#input-params-1
+
+        Returns:
+            dict: The parameters to create a litellm.SamplingParams in the model config.
+        """
+        args = {
+            "max_completion_tokens": self.max_new_tokens,
+            "stop": self.stop_tokens,
+            "temperature": self.temperature,
+            "top_p": self.top_p,
+            "seed": self.seed,
+            "repetition_penalty": self.repetition_penalty,
+            "frequency_penalty": self.frequency_penalty,
+        }
+        return {k: v for k, v in args.items() if v is not None}
+
     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