Use the programmatic interface using an already in memory loaded model

src/lighteval/logging/hierarchical_logger.py

@@ -34,8 +34,11 @@
 
     logger = get_logger(__name__, log_level="INFO")
 elif is_accelerate_available():
+    from accelerate import Accelerator, InitProcessGroupKwargs
     from accelerate.logging import get_logger
 
+    # We must init the accelerator before using the logger
+    accelerator = Accelerator(kwargs_handlers=[InitProcessGroupKwargs(timeout=timedelta(seconds=3000))])
     logger = get_logger(__name__, log_level="INFO")
 else:
     logger = Logger(__name__, level="INFO")

src/lighteval/models/adapter_model.py

@@ -41,7 +41,14 @@ class AdapterModel(BaseModel):
     def _create_auto_tokenizer(self, config: AdapterModelConfig, env_config: EnvConfig) -> PreTrainedTokenizer:
         # By default, we look at the model config for the model stored in `base_model`
         # (= the parent model, not the model of interest)
-        return self._create_auto_tokenizer_with_name(config.base_model, config=config, env_config=env_config)
+        return self._create_auto_tokenizer_with_name(
+            model_name=config.base_model,
+            revision=config.revision,
+            env_config=env_config,
+            tokenizer_name=config.tokenizer,
+            subfolder=config.subfolder,
+            trust_remote_code=config.trust_remote_code,
+        )
 
     def _create_auto_model(self, config: AdapterModelConfig, env_config: EnvConfig) -> AutoModelForCausalLM:
         """Returns a PeftModel from a base model and a version fined tuned using PEFT."""

src/lighteval/models/base_model.py

@@ -30,6 +30,7 @@
 from torch.utils.data import DataLoader
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
 
 from lighteval.data import GenerativeTaskDataset, LoglikelihoodDataset, LoglikelihoodSingleTokenDataset
 from lighteval.logging.hierarchical_logger import hlog, hlog_err, hlog_warn
@@ -57,6 +58,7 @@
 
 
 if is_accelerate_available():
+    from accelerate import Accelerator
     from accelerate.utils import calculate_maximum_sizes, convert_bytes, get_max_memory
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -67,8 +69,8 @@
 class BaseModel(LightevalModel):
     def __init__(
         self,
-        config: BaseModelConfig,
         env_config: EnvConfig,
+        config: BaseModelConfig,
     ):
         """Initializes a HuggingFace `AutoModel` and `AutoTokenizer` for evaluation."""
         self._config = config.init_configs(env_config)
@@ -114,6 +116,72 @@ def __init__(
 
         self.pairwise_tokenization = config.pairwise_tokenization
 
+    @classmethod
+    def from_model(
+        cls,
+        model: Union[AutoModelForCausalLM, LightevalModel],
+        env_config: EnvConfig,
+        accelerator: "Accelerator" = None,
+        tokenizer_name: str = None,  # custom tokenizer
+        trust_remote_code: bool = False,
+        use_chat_template: bool = False,
+        add_special_tokens: bool = True,
+        pairwise_tokenization: bool = False,
+        multichoice_continuations_start_space: bool = None,
+    ):
+        # Slightly hackish way to test if the model is a AutoModelForCausalLM, since the instances don't
+        # derive from this class explicitely
+        assert isinstance(model, LightevalModel) or type(model).__name__ in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values()
+
+        if isinstance(model, LightevalModel):
+            return model
+
+        # Instanciate the object without using __init__
+        self = cls.__new__(cls)
+        self._config = model.config
+        self._max_length = self._init_max_length(max_length=model.config.max_length)
+        self._tokenizer = self._create_auto_tokenizer_with_name(
+            model_name=model.name_or_path,
+            revision=model.config._commit_hash,
+            env_config=env_config,
+            trust_remote_code=trust_remote_code,
+            tokenizer_name=tokenizer_name,
+        )
+        self.model_name = _simplify_name(model.name_or_path)
+        self.model_sha = model.config._commit_hash
+
+        # If model_parallel is not set we compare the number of processes with the number of GPUs
+        self.model = model
+        self.model.eval()
+        torch.set_grad_enabled(False)
+
+        self.accelerator = accelerator
+        if accelerator is not None:
+            self._device = accelerator.device
+            self.model = self.accelerator.prepare(self.model.to(accelerator.device))
+        else:
+            self._device = "cpu"
+
+        self.use_chat_template = use_chat_template
+        self._add_special_tokens = add_special_tokens if add_special_tokens is not None else False
+        self.pairwise_tokenization = pairwise_tokenization
+        self.multichoice_continuations_start_space = multichoice_continuations_start_space
+
+        self.precision = _get_dtype(model.dtype, config=self._config)
+
+        if is_accelerate_available():
+            model_size, _ = calculate_maximum_sizes(self.model)
+            model_size = convert_bytes(model_size)
+        else:
+            model_size = -1
+        self.model_info = ModelInfo(
+            model_name=self.model_name,
+            model_sha=self.model_sha,
+            model_dtype=self.precision,
+            model_size=model_size,
+        )
+        return self
+
     @property
     def tokenizer(self):
         return self._tokenizer
@@ -207,10 +275,23 @@ def _create_auto_model(self, config: BaseModelConfig, env_config: EnvConfig) ->
     def _create_auto_tokenizer(
         self, config: BaseModelConfig, env_config: EnvConfig
     ) -> transformers.PreTrainedTokenizer:
-        return self._create_auto_tokenizer_with_name(config.pretrained, config=config, env_config=env_config)
+        return self._create_auto_tokenizer_with_name(
+            model_name=config.pretrained,
+            revision=config.revision,
+            env_config=env_config,
+            tokenizer_name=config.tokenizer,
+            subfolder=config.subfolder,
+            trust_remote_code=config.trust_remote_code,
+        )
 
     def _create_auto_tokenizer_with_name(
-        self, model_name: str, config: BaseModelConfig, env_config: EnvConfig
+        self,
+        model_name: str,
+        revision: str,
+        env_config: EnvConfig,
+        tokenizer_name: str = None,
+        subfolder: str = None,
+        trust_remote_code: bool = False,
     ) -> transformers.PreTrainedTokenizer:
         """
         Create a Hugging Face AutoTokenizer for language model.
@@ -231,25 +312,35 @@ def _create_auto_tokenizer_with_name(
         """
         try:
             tokenizer = AutoTokenizer.from_pretrained(
-                model_name if config.tokenizer is None else config.tokenizer,
-                revision=config.revision + (f"/{config.subfolder}" if config.subfolder is not None else ""),
+                model_name if tokenizer_name is None else tokenizer_name,
+                revision=revision + (f"/{subfolder}" if subfolder is not None else ""),
                 cache_dir=env_config.cache_dir,
                 token=env_config.token,
-                trust_remote_code=config.trust_remote_code,
+                trust_remote_code=trust_remote_code,
                 padding_side="left",
                 truncation_side="left",
             )
         except RecursionError:
             tokenizer = AutoTokenizer.from_pretrained(
-                model_name if config.tokenizer is None else config.tokenizer,
-                revision=config.revision + (f"/{config.subfolder}" if config.subfolder is not None else ""),
+                model_name if tokenizer_name is None else tokenizer_name,
+                revision=revision + (f"/{subfolder}" if subfolder is not None else ""),
                 cache_dir=env_config.cache_dir,
                 token=env_config.token,
-                trust_remote_code=config.trust_remote_code,
+                trust_remote_code=trust_remote_code,
                 unk_token="<unk>",
                 padding_side="left",
                 truncation_side="left",
             )
+        except FileNotFoundError:
+            hlog_warn("Problem when loading the tokenizer in the cache - discarding the provided cache path value.")
+            tokenizer = AutoTokenizer.from_pretrained(
+                model_name if tokenizer_name is None else tokenizer_name,
+                revision=revision + (f"/{subfolder}" if subfolder is not None else ""),
+                token=env_config.token,
+                trust_remote_code=trust_remote_code,
+                padding_side="left",
+                truncation_side="left",
+            )
         tokenizer.pad_token = tokenizer.eos_token
         tokenizer.model_max_length = self.max_length
         hlog("Tokenizer truncation and padding size set to the left side.")

src/lighteval/pipeline.py

@@ -34,7 +34,7 @@
 from lighteval.logging.evaluation_tracker import EvaluationTracker
 from lighteval.logging.hierarchical_logger import hlog, htrack_block
 from lighteval.metrics.utils.metric_utils import MetricCategory
-from lighteval.models.model_loader import load_model
+from lighteval.models.model_loader import BaseModel, load_model
 from lighteval.models.model_output import ModelResponse
 from lighteval.tasks.lighteval_task import LightevalTask, create_requests_from_tasks
 from lighteval.tasks.registry import Registry, taskinfo_selector
@@ -164,7 +164,15 @@ def _init_model(self, model_config, model):
                     )
                 else:
                     return load_model(config=model_config, env_config=self.pipeline_parameters.env_config)
-            return model
+            if isinstance(model, BaseModel):
+                return model
+            else:
+                return BaseModel.from_model(
+                    model=model,
+                    use_chat_template=self.pipeline_parameters.use_chat_template,
+                    env_config=self.pipeline_parameters.env_config,
+                    accelerator=self.accelerator,
+                )
 
     def _init_tasks_and_requests(self, tasks: str):
         with htrack_block("Tasks loading"):