FIX: Avoid needless copy from modules_to_save (huggingface#2220)

Resolves huggingface#2206

The problem is that we keep a "global" modules_to_save on the model
which contains all possible modules_to_save for each adapter. When the
first adapter targets layer "foo" with modules_to_save and the second
adapter targets "bar", then "foo" will create a copy of the original
module for the second adapter, even though it's not needed.

This does not change the result but is unnecessary and takes up memory.
Thus it should be avoided.

Author: BenjaminBossan

src/peft/mixed_model.py

@@ -251,7 +251,7 @@ def set_modules_to_save(self, peft_config: PeftConfig, adapter_name: str) -> Non
             self.modules_to_save = set(modules_to_save)
         else:
             self.modules_to_save.update(modules_to_save)
-        _set_trainable(self, adapter_name)
+        _set_trainable(self, adapter_name, modules_to_save=peft_config.modules_to_save)
 
     def set_adapter(self, adapter_name: Union[str, list[str]]) -> None:
         """

src/peft/peft_model.py

@@ -949,7 +949,8 @@ def set_additional_trainable_modules(self, peft_config, adapter_name):
                 self.modules_to_save = set(peft_config.modules_to_save)
             else:
                 self.modules_to_save.update(peft_config.modules_to_save)
-            _set_trainable(self, adapter_name)  # this may add a new ModulesToSaveWrapper
+            # this may add a new ModulesToSaveWrapper
+            _set_trainable(self, adapter_name, modules_to_save=peft_config.modules_to_save)
 
     def get_layer_status(self) -> list[TunerLayerStatus]:
         """Get the status of each adapter layer in the model.
@@ -1446,7 +1447,7 @@ def __init__(
                 break
 
         # to make sure classifier layer is trainable; this may add a new ModulesToSaveWrapper
-        _set_trainable(self, adapter_name)
+        _set_trainable(self, adapter_name, modules_to_save=peft_config.modules_to_save)
 
     def add_adapter(self, adapter_name: str, peft_config: PeftConfig, low_cpu_mem_usage: bool = False) -> None:
         """
@@ -2237,7 +2238,7 @@ def __init__(
                 break
 
         # to make sure classifier layer is trainable; this may add a new ModulesToSaveWrapper
-        _set_trainable(self, adapter_name)
+        _set_trainable(self, adapter_name, modules_to_save=peft_config.modules_to_save)
 
     def add_adapter(self, adapter_name: str, peft_config: PeftConfig, low_cpu_mem_usage: bool = False) -> None:
         """
@@ -2458,7 +2459,7 @@ def __init__(
                 break
 
         # to make sure classifier layer is trainable; this may add a new ModulesToSaveWrapper
-        _set_trainable(self, adapter_name)
+        _set_trainable(self, adapter_name, modules_to_save=peft_config.modules_to_save)
 
     def add_adapter(self, adapter_name: str, peft_config: PeftConfig, low_cpu_mem_usage: bool = False) -> None:
         """

src/peft/utils/other.py

@@ -275,8 +275,10 @@ def update(self, adapter_name):
 
                 context_manager = deepspeed.zero.GatheredParameters(self.original_module.parameters(), modifier_rank=0)
                 break
-        with context_manager:
-            self.modules_to_save.update(torch.nn.ModuleDict({adapter_name: copy.deepcopy(self.original_module)}))
+
+        if adapter_name not in self.modules_to_save:
+            with context_manager:
+                self.modules_to_save[adapter_name] = copy.deepcopy(self.original_module)
 
         if hasattr(self.modules_to_save[adapter_name], "_hf_hook"):
             old_hook = self.modules_to_save[adapter_name]._hf_hook
@@ -416,10 +418,10 @@ def _freeze_adapter(model, adapter_name):
             p.requires_grad = False
 
 
-def _set_trainable(model, adapter_name):
+def _set_trainable(model, adapter_name, modules_to_save):
     key_list = [key for key, _ in model.named_modules()]
     for key in key_list:
-        target_module_found = any(key.endswith(target_key) for target_key in model.modules_to_save)
+        target_module_found = any(key.endswith(target_key) for target_key in modules_to_save)
         if target_module_found:
             parent, target, target_name = _get_submodules(model, key)
             if isinstance(target, ModulesToSaveWrapper):

tests/test_custom_models.py

@@ -1608,6 +1608,26 @@ def test_multiple_adapters_seq_cls_mixed_modules_to_save_merging_adapters(self):
         with pytest.raises(ValueError, match=msg):
             model.add_weighted_adapter(["default", "other"], weights=[1.0, 1.0], adapter_name="merged")
 
+    def test_multiple_adapters_no_needless_copy_modules_to_save(self):
+        # See 2206
+        # The problem was that we keep a "global" modules_to_save on the model which contains all possible
+        # modules_to_save for each adapter. When the first adapter targets embed_tokens with modules_to_save and the
+        # second adapter targets lm_head, then embed_tokens will create a copy of the original module for the second
+        # adapter, even though it's not needed. The copy still acts as expected but uses unnecessary memory.
+        model_id = "hf-internal-testing/tiny-random-OPTForCausalLM"
+        model = AutoModelForCausalLM.from_pretrained(model_id).to(self.torch_device)
+        config0 = LoraConfig(modules_to_save=["embed_tokens"])
+        config1 = LoraConfig(modules_to_save=["lm_head"])
+        model = get_peft_model(model, config0)
+        model.add_adapter("other", config1)
+
+        lm_head_keys = list(model.base_model.model.lm_head.modules_to_save.keys())
+        assert lm_head_keys == ["other"]
+
+        embed_token_keys = list(model.base_model.model.model.decoder.embed_tokens.modules_to_save.keys())
+        # before the fix, this would be: ['default', 'other']
+        assert embed_token_keys == ["default"]
+
     def test_existing_model_card(self):
         # ensure that if there is already a model card, it is not overwritten
         model = MLP()