Allow AdapterModels to have custom tokens (#306)

---------

Co-authored-by: Clémentine Fourrier <[email protected]>

Author: mapmeld

src/lighteval/main_accelerate.py

@@ -169,11 +169,11 @@ def accelerate(  # noqa C901
         # Keeping only non null params
         args_dict = {k: v for k, v in args_dict.items() if v is not None}
 
-        if config["merged_weights"]["delta_weights"]:
+        if config["merged_weights"].get("delta_weights", False):
             if config["merged_weights"]["base_model"] is None:
                 raise ValueError("You need to specify a base model when using delta weights")
             model_config = DeltaModelConfig(**args_dict)
-        elif config["merged_weights"]["adapter_weights"]:
+        elif config["merged_weights"].get("adapter_weights", False):
             if config["merged_weights"]["base_model"] is None:
                 raise ValueError("You need to specify a base model when using adapter weights")
             model_config = AdapterModelConfig(**args_dict)

src/lighteval/models/transformers/adapter_model.py

@@ -84,6 +84,18 @@ def _create_auto_model(self, config: AdapterModelConfig, env_config: EnvConfig)
             base = AutoModelForCausalLM.from_pretrained(
                 config.base_model, torch_dtype=torch.float16, low_cpu_mem_usage=True, token=env_config.token
             )
+            # resize model for adapters with added tokens
+            token_diff = len(self._tokenizer) - base.config.vocab_size
+            if token_diff != 0:
+                if token_diff > 0:
+                    logger.info(
+                        f"You're using the adapter model's tokenizer, which has more tokens than the base model. Adding {token_diff} token(s)."
+                    )
+                else:
+                    logger.info(
+                        f"You're using the adapter model's tokenizer, which has fewer tokens than the base model. Removing {abs(token_diff)} token(s)."
+                    )
+                base.resize_token_embeddings(len(self._tokenizer))
             # Should pass revision
             model = PeftModel.from_pretrained(base, adapter_weights)
             model = model.merge_and_unload()