Homogeneize generation params (huggingface#428)

This PR does 3 things:

    Provide an homogeneized API for people to use to provide model generation parameters in model configs. Those parameters are notably provided to all models which can take them (vllm, open ai, tgi, transformers, ...)
    Renames BaseModel to TransformersModel
    Also allows TransformersModels to use a transformers.GenerationConfig object directly, when created programmatically

I would put system_prompt, fewshot_seeds, and use_chat_template in the GenerationParameters too since they are generation parameters logically, but it can be another PR

---------

Co-authored-by: Nathan Habib <30601243+NathanHB@users.noreply.github.com>
Co-authored-by: Albert Villanova del Moral <8515462+albertvillanova@users.noreply.github.com>

Author: 3 people

docs/source/package_reference/models.mdx

@@ -6,9 +6,9 @@
 
 
 ## Accelerate and Transformers Models
-### BaseModel
-[[autodoc]] models.transformers.base_model.BaseModelConfig
-[[autodoc]] models.transformers.base_model.BaseModel
+### TransformersModel
+[[autodoc]] models.transformers.transformers_model.TransformersModelConfig
+[[autodoc]] models.transformers.transformers_model.TransformersModel
 
 ### AdapterModel
 [[autodoc]] models.transformers.adapter_model.AdapterModelConfig

examples/model_configs/base_model.yaml renamed to examples/model_configs/transformers_model.yaml

@@ -1,6 +1,6 @@
 model:
   base_params:
-    model_args: "pretrained=HuggingFaceH4/zephyr-7b-beta,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
+    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
     dtype: "bfloat16"
     compile: true
   merged_weights: # Ignore this section if you are not using PEFT models
@@ -9,3 +9,4 @@ model:
     base_model: null # path to the base_model
   generation:
     multichoice_continuations_start_space: null # If true/false, will force multiple choice continuations to start/not start with a space. If none, will do nothing
+    temperature: 0.5

src/lighteval/__main__.py

@@ -20,7 +20,7 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 import logging
-from logging.config import dictConfig
+import logging.config
 
 import colorlog
 import typer
@@ -57,7 +57,8 @@
     },
 )
 
-dictConfig(logging_config)
+logging.config.dictConfig(logging_config)
+logging.captureWarnings(capture=True)
 
 app.command(rich_help_panel="Evaluation Backends")(lighteval.main_accelerate.accelerate)
 app.command(rich_help_panel="Evaluation Utils")(lighteval.main_baseline.baseline)

src/lighteval/main_accelerate.py

@@ -44,7 +44,7 @@ def accelerate(  # noqa C901
     model_args: Annotated[
         str,
         Argument(
-            help="Model arguments in the form key1=value1,key2=value2,... or path to yaml config file (see examples/model_configs/base_model.yaml)"
+            help="Model arguments in the form key1=value1,key2=value2,... or path to yaml config file (see examples/model_configs/transformers_model.yaml)"
         ),
     ],
     tasks: Annotated[str, Argument(help="Comma-separated list of tasks to evaluate on.")],
@@ -107,9 +107,10 @@ def accelerate(  # noqa C901
     from accelerate import Accelerator, InitProcessGroupKwargs
 
     from lighteval.logging.evaluation_tracker import EvaluationTracker
+    from lighteval.models.model_input import GenerationParameters
     from lighteval.models.transformers.adapter_model import AdapterModelConfig
-    from lighteval.models.transformers.base_model import BaseModelConfig, BitsAndBytesConfig
     from lighteval.models.transformers.delta_model import DeltaModelConfig
+    from lighteval.models.transformers.transformers_model import BitsAndBytesConfig, TransformersModelConfig
     from lighteval.pipeline import EnvConfig, ParallelismManager, Pipeline, PipelineParameters
 
     accelerator = Accelerator(kwargs_handlers=[InitProcessGroupKwargs(timeout=timedelta(seconds=3000))])
@@ -154,6 +155,8 @@ def accelerate(  # noqa C901
         # We extract the model args
         args_dict = {k.split("=")[0]: k.split("=")[1] for k in config["base_params"]["model_args"].split(",")}
 
+        args_dict["generation_parameters"] = GenerationParameters.from_dict(config)
+
         # We store the relevant other args
         args_dict["base_model"] = config["merged_weights"]["base_model"]
         args_dict["compile"] = bool(config["base_params"]["compile"])
@@ -180,13 +183,13 @@ def accelerate(  # noqa C901
         elif config["merged_weights"]["base_model"] not in ["", None]:
             raise ValueError("You can't specify a base model if you are not using delta/adapter weights")
         else:
-            model_config = BaseModelConfig(**args_dict)
+            model_config = TransformersModelConfig(**args_dict)
     else:
         model_args_dict: dict = {k.split("=")[0]: k.split("=")[1] if "=" in k else True for k in model_args.split(",")}
         model_args_dict["accelerator"] = accelerator
         model_args_dict["use_chat_template"] = use_chat_template
         model_args_dict["compile"] = bool(model_args_dict["compile"]) if "compile" in model_args_dict else False
-        model_config = BaseModelConfig(**model_args_dict)
+        model_config = TransformersModelConfig(**model_args_dict)
 
     pipeline = Pipeline(
         tasks=tasks,

src/lighteval/main_endpoint.py

@@ -42,8 +42,11 @@
 @app.command(rich_help_panel="Evaluation Backends")
 def openai(
     # === general ===
-    model_name: Annotated[
-        str, Argument(help="The model name to evaluate (has to be available through the openai API.")
+    model_args: Annotated[
+        str,
+        Argument(
+            help="Model name as a string (has to be available through the openai API) or path to yaml config file (see examples/model_configs/transformers_model.yaml)"
+        ),
     ],
     tasks: Annotated[str, Argument(help="Comma-separated list of tasks to evaluate on.")],
     # === Common parameters ===
@@ -96,6 +99,11 @@ def openai(
     from lighteval.models.endpoints.openai_model import OpenAIModelConfig
     from lighteval.pipeline import EnvConfig, ParallelismManager, Pipeline, PipelineParameters
 
+    if model_args.endswith(".yaml"):
+        model_config = OpenAIModelConfig.from_path(model_args)
+    else:
+        model_config = OpenAIModelConfig(model=model_args)
+
     env_config = EnvConfig(token=TOKEN, cache_dir=cache_dir)
     evaluation_tracker = EvaluationTracker(
         output_dir=output_dir,
@@ -107,7 +115,6 @@ def openai(
     )
 
     parallelism_manager = ParallelismManager.OPENAI
-    model_config = OpenAIModelConfig(model=model_name)
 
     pipeline_params = PipelineParameters(
         launcher_type=parallelism_manager,
@@ -205,7 +212,6 @@ def inference_endpoint(
     """
     Evaluate models using inference-endpoints as backend.
     """
-
     from lighteval.logging.evaluation_tracker import EvaluationTracker
     from lighteval.models.endpoints.endpoint_model import InferenceEndpointModelConfig, ServerlessEndpointModelConfig
     from lighteval.pipeline import EnvConfig, ParallelismManager, Pipeline, PipelineParameters
@@ -319,7 +325,6 @@ def tgi(
     """
     Evaluate models using TGI as backend.
     """
-
     from lighteval.logging.evaluation_tracker import EvaluationTracker
     from lighteval.models.endpoints.tgi_model import TGIModelConfig
     from lighteval.pipeline import EnvConfig, ParallelismManager, Pipeline, PipelineParameters

src/lighteval/main_vllm.py

@@ -37,7 +37,12 @@
 
 def vllm(
     # === general ===
-    model_args: Annotated[str, Argument(help="Model arguments in the form key1=value1,key2=value2,...")],
+    model_args: Annotated[
+        str,
+        Argument(
+            help="Model arguments in the form key1=value1,key2=value2,... or path to yaml config file (see examples/model_configs/transformers_model.yaml)"
+        ),
+    ],
     tasks: Annotated[str, Argument(help="Comma-separated list of tasks to evaluate on.")],
     # === Common parameters ===
     use_chat_template: Annotated[
@@ -88,7 +93,10 @@ def vllm(
     """
     Evaluate models using vllm as backend.
     """
+    import yaml
+
     from lighteval.logging.evaluation_tracker import EvaluationTracker
+    from lighteval.models.model_input import GenerationParameters
     from lighteval.models.vllm.vllm_model import VLLMModelConfig
     from lighteval.pipeline import EnvConfig, ParallelismManager, Pipeline, PipelineParameters
 
@@ -118,8 +126,15 @@ def vllm(
         system_prompt=system_prompt,
     )
 
-    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)
+    if model_args.endswith(".yaml"):
+        with open(model_args, "r") as f:
+            config = yaml.safe_load(f)["model"]
+        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)
 
     pipeline = Pipeline(
         tasks=tasks,

src/lighteval/models/endpoints/endpoint_model.py

@@ -24,7 +24,7 @@
 import logging
 import re
 import time
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 from typing import Coroutine, Dict, List, Optional, Union
 
 import requests
@@ -35,6 +35,7 @@
     InferenceEndpoint,
     InferenceEndpointError,
     InferenceEndpointTimeoutError,
+    TextGenerationInputGenerateParameters,
     TextGenerationInputGrammarType,
     TextGenerationOutput,
     create_inference_endpoint,
@@ -48,6 +49,7 @@
 
 from lighteval.data import GenerativeTaskDataset, LoglikelihoodDataset
 from lighteval.models.abstract_model import LightevalModel, ModelInfo
+from lighteval.models.model_input import GenerationParameters
 from lighteval.models.model_output import GenerativeResponse, LoglikelihoodResponse, LoglikelihoodSingleTokenResponse
 from lighteval.tasks.requests import (
     GreedyUntilRequest,
@@ -78,6 +80,11 @@
 class ServerlessEndpointModelConfig:
     model_name: str
     add_special_tokens: bool = True
+    generation_parameters: GenerationParameters = None
+
+    def __post_init__(self):
+        if not self.generation_parameters:
+            self.generation_parameters = GenerationParameters()
 
     @classmethod
     def from_path(cls, path: str) -> "ServerlessEndpointModelConfig":
@@ -106,6 +113,7 @@ class InferenceEndpointModelConfig:
     namespace: str = None  # The namespace under which to launch the endpoint. Defaults to the current user's namespace
     image_url: str = None
     env_vars: dict = None
+    generation_parameters: GenerationParameters = None
 
     def __post_init__(self):
         # xor operator, one is None but not the other
@@ -117,6 +125,9 @@ def __post_init__(self):
         if not (self.endpoint_name is None) ^ int(self.model_name is None):
             raise ValueError("You need to set either endpoint_name or model_name (but not both).")
 
+        if not self.generation_parameters:
+            self.generation_parameters = GenerationParameters()
+
     @classmethod
     def from_path(cls, path: str) -> "InferenceEndpointModelConfig":
         """Load configuration for inference endpoint model from YAML file path.
@@ -305,6 +316,8 @@ def __init__(  # noqa: C901
             model_dtype=getattr(config, "model_dtype", "default"),
             model_size=-1,
         )
+        self.generation_parameters = config.generation_parameters
+        self.generation_config = TextGenerationInputGenerateParameters(**self.generation_parameters.to_tgi_ie_dict())
 
     @staticmethod
     def get_larger_hardware_suggestion(cur_instance_type: str = None, cur_instance_size: str = None):
@@ -388,16 +401,17 @@ def _async_process_request(
     ) -> Coroutine[None, list[TextGenerationOutput], str]:
         # Todo: add an option to launch with conversational instead for chat prompts
         # https://huggingface.co/docs/huggingface_hub/v0.20.3/en/package_reference/inference_client#huggingface_hub.AsyncInferenceClient.conversational
-        generated_text = self.async_client.text_generation(
-            prompt=context,
+        generation_config: TextGenerationInputGenerateParameters = replace(
+            self.generation_config,
+            stop=stop_tokens,
+            max_new_tokens=max_tokens,
             details=True,
             decoder_input_details=True,
             grammar=grammar,
-            max_new_tokens=max_tokens,
-            stop_sequences=stop_tokens,
-            # truncate=,
         )
 
+        generated_text = self.async_client.text_generation(prompt=context, generation_config=generation_config)
+
         return generated_text
 
     def _process_request(
@@ -409,14 +423,18 @@ def _process_request(
     ) -> TextGenerationOutput:
         # Todo: add an option to launch with conversational instead for chat prompts
         # https://huggingface.co/docs/huggingface_hub/v0.20.3/en/package_reference/inference_client#huggingface_hub.AsyncInferenceClient.conversational
-        generated_text = self.client.text_generation(
-            prompt=context,
+        generation_config: TextGenerationInputGenerateParameters = replace(
+            self.generation_config,
+            stop=stop_tokens,
+            max_new_tokens=max_tokens,
             details=True,
             decoder_input_details=True,
             grammar=grammar,
-            max_new_tokens=max_tokens,
-            stop_sequences=stop_tokens,
-            # truncate=,
+        )
+
+        generated_text = self.client.text_generation(
+            prompt=context,
+            generation_config=generation_config,
         )
 
         return generated_text

src/lighteval/models/endpoints/openai_model.py

@@ -32,6 +32,7 @@
 from lighteval.data import GenerativeTaskDataset, LoglikelihoodDataset
 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,
@@ -62,14 +63,30 @@
 @dataclass
 class OpenAIModelConfig:
     model: str
+    generation_parameters: GenerationParameters = None
+
+    def __post_init__(self):
+        if not self.generation_parameters:
+            self.generation_parameters = GenerationParameters()
+
+    @classmethod
+    def from_path(cls, path: str) -> "OpenAIModelConfig":
+        import yaml
+
+        with open(path, "r") as f:
+            config = yaml.safe_load(f)["model"]
+        generation_parameters = GenerationParameters.from_dict(config)
+        return cls(model=config["model_name"], generation_parameters=generation_parameters)
 
 
 class OpenAIClient(LightevalModel):
     _DEFAULT_MAX_LENGTH: int = 4096
 
-    def __init__(self, config, env_config) -> None:
+    def __init__(self, config: OpenAIModelConfig, env_config) -> None:
         api_key = os.environ["OPENAI_API_KEY"]
         self.client = OpenAI(api_key=api_key)
+        self.generation_parameters = config.generation_parameters
+        self.sampling_params = self.generation_parameters.to_vllm_openai_dict()
 
         self.model_info = ModelInfo(
             model_name=config.model,
@@ -96,6 +113,7 @@ def __call_api(self, prompt, return_logits, max_new_tokens, num_samples, logit_b
                     logprobs=return_logits,
                     logit_bias=logit_bias,
                     n=num_samples,
+                    **self.sampling_params,
                 )
                 return response
             except Exception as e: