diff --git a/tests/test_reward_trainer.py b/tests/test_reward_trainer.py
index 51ea1183f7..5a858b2f38 100644
--- a/tests/test_reward_trainer.py
+++ b/tests/test_reward_trainer.py
@@ -189,6 +189,28 @@ def test_train_lora_pretokenized(self):
                 new_param = trainer.model.get_parameter(n)
                 self.assertTrue(torch.allclose(param, new_param, atol=1e-12, rtol=1e-12))
 
+    @require_peft
+    def test_all_linear_user_warning(self):
+        peft_config = LoraConfig(
+            task_type=TaskType.SEQ_CLS,
+            inference_mode=False,
+            r=8,
+            lora_alpha=32,
+            lora_dropout=0.1,
+            target_modules="all-linear",
+        )
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            dummy_dataset = load_dataset("trl-internal-testing/zen", "conversational_preference", split="train")
+            training_args = RewardConfig(output_dir=tmp_dir, max_steps=3, report_to="none")
+            with self.assertWarns(UserWarning, msg="You passed target_modules="):
+                _ = RewardTrainer(
+                    model=self.model,
+                    args=training_args,
+                    processing_class=self.tokenizer,
+                    train_dataset=dummy_dataset,
+                    peft_config=peft_config,
+                )
+
     def test_margin(self):
         with tempfile.TemporaryDirectory() as tmp_dir:
             dummy_dataset_dict = {
diff --git a/trl/trainer/reward_trainer.py b/trl/trainer/reward_trainer.py
index 063fe5e8e8..6069104608 100644
--- a/trl/trainer/reward_trainer.py
+++ b/trl/trainer/reward_trainer.py
@@ -25,6 +25,7 @@
 from accelerate import PartialState
 from accelerate.utils import gather_object
 from datasets import Dataset
+from peft.utils import INCLUDE_LINEAR_LAYERS_SHORTHAND
 from transformers import (
     BaseImageProcessor,
     DataCollator,
@@ -159,6 +160,19 @@ def __init__(
 
                     model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)
 
+                # Warn if the user passes "all-linear" for the target_modules
+                target_modules = (
+                    peft_config.get("target_modules", None)
+                    if isinstance(peft_config, dict)
+                    else peft_config.target_modules
+                )
+                if target_modules == INCLUDE_LINEAR_LAYERS_SHORTHAND:
+                    warnings.warn(
+                        f"You passed target_modules='{INCLUDE_LINEAR_LAYERS_SHORTHAND}' in the peft_config. "
+                        "This will result in all linear layers except the output layer being adapted. "
+                        "This could negatively impact the performance of the reward model as the output layer (used for scoring of chosen and rejected completions) will not be adapted or trained.",
+                        UserWarning,
+                    )
                 model = get_peft_model(model, peft_config)
 
         # Disable dropout in the model