yo

llmfoundry/models/hf/hf_base.py

@@ -72,7 +72,6 @@ def __init__(
         additional_train_metrics: Optional[list] = None,
         additional_eval_metrics: Optional[list] = None,
         should_save_peft_only: bool = True,
-        peft_is_trainable: bool = False,
     ):
         config_overrides = config_overrides or {}
 
@@ -86,7 +85,6 @@ def __init__(
             config_overrides=config_overrides,
             load_in_8bit=load_in_8bit,
             pretrained=pretrained,
-            peft_is_trainable=peft_is_trainable,
         )
 
         model = self.transform_model(model)
@@ -209,7 +207,6 @@ def build_inner_model(
         pretrained: bool,
         model_cls: Optional[Union[_BaseAutoModelClass, PreTrainedModel]] = None,
         prepare_for_fsdp: bool = False,
-        peft_is_trainable: bool = False,
     ) -> Union[PreTrainedModel, 'PeftModel']:
         """Builds the inner model for the ComposerHFCausalLM.
 
@@ -225,7 +222,6 @@ def build_inner_model(
             pretrained (bool): Whether the model is pretrained.
             model_cls (Union[Type, Type[PreTrainedModel]]): Kept for backwards compatibility.
             prepare_for_fsdp (bool, optional): Kept for backwards compatilbility.
-            peft_is_trainable (bool): Whether loaded PEFT adapters are trainable. Default: ``False``.
 
         Returns:
             Union[PreTrainedModel, 'PeftModel']: The built inner model.
@@ -396,7 +392,7 @@ def build_inner_model(
             model = PeftModelForCausalLM.from_pretrained(
                 model,
                 pretrained_lora_id_or_path,
-                is_trainable=peft_is_trainable,
+                is_trainable=True,
             )
 
         if prepare_for_fsdp:

llmfoundry/models/hf/hf_causal_lm.py

@@ -54,7 +54,6 @@ class ComposerHFCausalLM(BaseHuggingFaceModel):
         init_device (str, optional): Which device to initialize the model on. Default: ``'cpu'``.
         use_flash_attention_2 (bool, optional): Whether to use flash-attention 2. Default: ``False``.
         tokenizer (PreTrainedTokenizer): The tokenizer that the model will use.
-        peft_is_trainable (bool): Whether loaded PEFT adapters are trainable. Default: ``False``.
     """
 
     model_cls: Union[_BaseAutoModelClass,
@@ -80,7 +79,6 @@ def __init__(
         additional_train_metrics: Optional[list] = None,
         additional_eval_metrics: Optional[list] = None,
         should_save_peft_only: bool = True,
-        peft_is_trainable: bool = False,
     ):
         super().__init__(
             pretrained_model_name_or_path,
@@ -100,5 +98,4 @@ def __init__(
             additional_train_metrics=additional_train_metrics,
             additional_eval_metrics=additional_eval_metrics,
             should_save_peft_only=should_save_peft_only,
-            peft_is_trainable=peft_is_trainable,
         )

tests/models/hf/test_hf_base.py

@@ -27,8 +27,7 @@ def test_build_inner_model_fsdp():
     assert model.fsdp_wrap_fn(model.model.layers[0])
 
 
-@pytest.mark.parametrize('trainable', [True, False])
-def test_pretrained_peft_trainable(trainable: bool):
+def test_pretrained_peft_trainable():
     model = BaseHuggingFaceModel.build_inner_model(
         pretrained_model_name_or_path='facebook/opt-350m',
         pretrained_lora_id_or_path='ybelkada/opt-350m-lora',
@@ -40,15 +39,10 @@ def test_pretrained_peft_trainable(trainable: bool):
         load_in_8bit=False,
         pretrained=True,
         prepare_for_fsdp=True,
-        peft_is_trainable=trainable,
     )
 
     assert isinstance(model, PeftModel)
 
     n_trainable, n_all = model.get_nb_trainable_parameters()
     assert n_all > 0
-
-    if trainable:
-        assert n_trainable > 0
-    else:
-        assert n_trainable == 0
+    assert n_trainable > 0