Fix missing PEFT availability check when passing peft_config to experimental trainers

trl/experimental/bco/bco_trainer.py

@@ -65,7 +65,7 @@
 
 
 if is_peft_available():
-    from peft import PeftModel, get_peft_model, prepare_model_for_kbit_training
+    from peft import PeftConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
 
 if is_wandb_available():
     import wandb
@@ -469,15 +469,21 @@ def __init__(
         if isinstance(ref_model, str):
             ref_model = AutoModelForCausalLM.from_pretrained(ref_model, **model_init_kwargs)
 
+        # PEFT
         # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
         # has been called in order to properly call autocast if needed.
         self._peft_has_been_casted_to_bf16 = False
-
-        if not is_peft_available() and peft_config is not None:
-            raise ValueError(
-                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it with `pip install peft` to use the PEFT models"
-            )
-        elif is_peft_available() and peft_config is not None:
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
             if isinstance(model, PeftModel):
                 raise ValueError(
                     "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first "

trl/experimental/cpo/cpo_trainer.py

@@ -61,7 +61,7 @@
 
 
 if is_peft_available():
-    from peft import PeftModel, get_peft_model, prepare_model_for_kbit_training
+    from peft import PeftConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
 
 
 if is_wandb_available():
@@ -169,15 +169,21 @@ def __init__(
         if isinstance(model, str):
             model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)
 
+        # PEFT
         # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
         # has been called in order to properly call autocast if needed.
         self._peft_has_been_casted_to_bf16 = False
-
-        if not is_peft_available() and peft_config is not None:
-            raise ValueError(
-                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
-            )
-        elif is_peft_available() and peft_config is not None:
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
             if isinstance(model, PeftModel):
                 raise ValueError(
                     "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first "

trl/experimental/distillation/distillation_trainer.py

@@ -430,6 +430,16 @@ def __init__(
 
         # ── PEFT ──
         if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
             model = get_peft_model(model, peft_config)
 
         # ── Data collator ──

trl/experimental/kto/kto_trainer.py

@@ -67,7 +67,7 @@
     from liger_kernel.chunked_loss import LigerFusedLinearKTOLoss
 
 if is_peft_available():
-    from peft import PeftModel, get_peft_model, prepare_model_for_kbit_training
+    from peft import PeftConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
 
 
 if TYPE_CHECKING:
@@ -282,20 +282,27 @@ def __init__(
         if tokenizer.pad_token is None:
             tokenizer.pad_token = tokenizer.eos_token
 
+        # PEFT
         # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
         # has been called in order to properly call autocast if needed.
         self._peft_has_been_casted_to_bf16 = False
-
-        if not is_peft_available() and peft_config is not None:
-            raise ValueError(
-                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it with `pip install peft` to use the PEFT models"
-            )
-        if is_peft_available() and isinstance(model, PeftModel) and peft_config is not None:
-            raise ValueError(
-                "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first merge "
-                "and unload the existing adapter, save the resulting base model, and then pass that base model along "
-                "with the new `peft_config` to the trainer."
-            )
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
+            if isinstance(model, PeftModel):
+                raise ValueError(
+                    "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first merge "
+                    "and unload the existing adapter, save the resulting base model, and then pass that base model along "
+                    "with the new `peft_config` to the trainer."
+                )
         if is_peft_available() and isinstance(model, PeftModel) and ref_model is None:
             # If the model is a PEFT model with a pretrained adapter, we need to create a "ref" adapter that is a copy
             # of the "default" adapter, so that we can use it as the reference model during KTO training.

trl/experimental/online_dpo/online_dpo_trainer.py

@@ -282,6 +282,18 @@ def __init__(
         self.is_encoder_decoder = model.config.is_encoder_decoder
         self.is_vision_model = model.config.model_type in MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES.keys()
 
+        # PEFT
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
         if peft_config is not None or (is_peft_available() and isinstance(model, PeftModel)):
             model = prepare_peft_model(model, peft_config, args)
 

trl/experimental/orpo/orpo_trainer.py

@@ -62,7 +62,7 @@
 
 
 if is_peft_available():
-    from peft import PeftModel, get_peft_model, prepare_model_for_kbit_training
+    from peft import PeftConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training
 
 
 if is_wandb_available():
@@ -178,15 +178,21 @@ def __init__(
         if isinstance(model, str):
             model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)
 
+        # PEFT
         # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
         # has been called in order to properly call autocast if needed.
         self._peft_has_been_casted_to_bf16 = False
-
-        if not is_peft_available() and peft_config is not None:
-            raise ValueError(
-                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
-            )
-        elif is_peft_available() and peft_config is not None:
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
             if isinstance(model, PeftModel):
                 raise ValueError(
                     "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first "

trl/experimental/ppo/ppo_trainer.py

@@ -416,12 +416,18 @@ def __init__(
                 "[Approximating KL Divergence](http://joschu.net/blog/kl-approx.html) for details."
             )
 
-        # peft support
-        if not is_peft_available() and peft_config is not None:
-            raise ImportError(
-                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
-            )
-        elif is_peft_available() and peft_config is not None:
+        # PEFT
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
             if isinstance(self.policy_model, PeftModel):
                 raise ValueError(
                     "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first "

trl/experimental/prm/prm_trainer.py

@@ -44,7 +44,7 @@
 
 
 if is_peft_available():
-    from peft import PeftModel
+    from peft import PeftConfig, PeftModel
 
 logger = logging.get_logger(__name__)
 
@@ -172,6 +172,18 @@ def __init__(
         if train_dataset is None:
             raise ValueError("`train_dataset` is required")
 
+        # PEFT
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
         if peft_config is not None or (is_peft_available() and isinstance(model, PeftModel)):
             model = prepare_peft_model(model, peft_config, args)
 

trl/experimental/tpo/tpo_trainer.py

@@ -350,6 +350,18 @@ def __init__(
         if tokenizer.pad_token is None:
             tokenizer.pad_token = tokenizer.eos_token
 
+        # PEFT
+        if peft_config is not None:
+            if not is_peft_available():
+                raise ImportError(
+                    "You passed `peft_config` but the `peft` library is not installed. "
+                    "Install it with `pip install trl[peft]`."
+                )
+            if not isinstance(peft_config, PeftConfig):
+                raise TypeError(
+                    f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
+                    f"got {type(peft_config).__name__}."
+                )
         if is_peft_available() and is_peft_model(model) and peft_config is not None:
             raise ValueError(
                 "You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first merge "