[WIP] ENH: Adapter injection based on state_dict

docs/source/developer_guides/low_level_api.md

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
 
 # Adapter injection
 
-With PEFT, you can inject trainable adapters into any `torch` module which allows you to use adapter methods without relying on the modeling classes in PEFT. Currently, PEFT supports injecting [LoRA](../conceptual_guides/adapter#low-rank-adaptation-lora), [AdaLoRA](../conceptual_guides/adapter#adaptive-low-rank-adaptation-adalora), and [IA3](../conceptual_guides/ia3) into models because for these adapters, inplace modification of the model is sufficient for finetuning it.
+With PEFT, you can inject trainable adapters into any `torch` module which allows you to use adapter methods without relying on the modeling classes in PEFT. This works for all adapters except for those based on prompt learning (e.g. prefix tuning or p-tuning).
 
 Check the table below to see when you should inject adapters.
 
@@ -87,6 +87,21 @@ DummyModel(
 )
 ```
 
+Sometimes, it is possible that there is a PEFT adapter checkpoint but the corresponding PEFT config is not known for whatever reason. To inject the PEFT layers for this checkpoint, you would usually have to reverse-engineer the corresponding PEFT config, most notably the `target_modules` argument, based on the `state_dict` from the checkpoint. This can be cumbersome and error prone. To avoid this, it is also possible to call [`inject_adapter_in_model`] and pass the loaded `state_dict` as an argument:
+
+```python
+from safetensors.torch import load_file
+
+model = ...
+state_dict = load_file(<path-to-safetensors-file>)
+lora_config = LoraConfig(...)
+model = inject_adapter_in_model(lora_config, model, state_dict=state_dict)
+```
+
+In this case, PEFT will use the `state_dict` as reference for which layers to target instead of using the PEFT config. As a user, you don't have to set the exact `target_modules` of the PEFT config for this to work. However, you should still pass a PEFT config of the right type, in this example `LoraConfig`, you can leave the `target_modules` as `None`. Be aware that this still only creates the uninitialized PEFT layers, the values from the `state_dict` are not used to populate the model weights.
+
+⚠️ Note that if there is a mismatch between what is configured in the PEFT config and what is found in the `state_dict`, PEFT will warn you about this. You can ignore the warning if you know that the PEFT config is not correctly specified.
+
 ## Saving the model
 
 To only save the adapter, use the [`get_peft_model_state_dict`] function:

src/peft/mapping.py

@@ -14,7 +14,7 @@
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, Optional
 
 import torch
 
@@ -45,7 +45,11 @@ def get_peft_config(config_dict: dict[str, Any]) -> PeftConfig:
 
 
 def inject_adapter_in_model(
-    peft_config: PeftConfig, model: torch.nn.Module, adapter_name: str = "default", low_cpu_mem_usage: bool = False
+    peft_config: PeftConfig,
+    model: torch.nn.Module,
+    adapter_name: str = "default",
+    low_cpu_mem_usage: bool = False,
+    state_dict: Optional[dict[str, torch.Tensor]] = None,
 ) -> torch.nn.Module:
     r"""
     A simple API to create and inject adapter in-place into a model. Currently the API does not support prompt learning
@@ -61,6 +65,11 @@ def inject_adapter_in_model(
             The name of the adapter to be injected, if not provided, the default adapter name is used ("default").
         low_cpu_mem_usage (`bool`, `optional`, defaults to `False`):
             Create empty adapter weights on meta device. Useful to speed up the loading process.
+        state_dict (`dict`, *optional*, defaults to `None`)
+            If a state_dict is passed here, the adapters will be injected based on the entries of the state_dict. This
+            can be useful when the exact `target_modules` of the PEFT method is unknown, for instance because the
+            checkpoint was created without meta data. Note that the values from the state_dict are not used, only the
+            keys are used to determine the correct layers that should be adapted.
     """
     if peft_config.is_prompt_learning or peft_config.is_adaption_prompt:
         raise ValueError("`create_and_replace` does not support prompt learning and adaption prompt yet.")
@@ -73,6 +82,8 @@ def inject_adapter_in_model(
     tuner_cls = PEFT_TYPE_TO_TUNER_MAPPING[peft_config.peft_type]
 
     # By instantiating a peft model we are injecting randomly initialized LoRA layers into the model's modules.
-    peft_model = tuner_cls(model, peft_config, adapter_name=adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
+    peft_model = tuner_cls(
+        model, peft_config, adapter_name=adapter_name, low_cpu_mem_usage=low_cpu_mem_usage, state_dict=state_dict
+    )
 
     return peft_model.model

src/peft/tuners/adalora/model.py

@@ -65,8 +65,8 @@ class AdaLoraModel(LoraModel):
 
     # Note: don't redefine prefix here, it should be inherited from LoraModel
 
-    def __init__(self, model, config, adapter_name):
-        super().__init__(model, config, adapter_name)
+    def __init__(self, model, config, adapter_name, **kwargs):
+        super().__init__(model, config, adapter_name, **kwargs)
 
         traininable_mode_counter = 0
         for config in self.peft_config.values():

src/peft/tuners/boft/model.py

@@ -74,9 +74,6 @@ class BOFTModel(BaseTuner):
 
     prefix: str = "boft_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: BOFTConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/c3a/model.py

@@ -55,9 +55,6 @@ class C3AModel(BaseTuner):
 
     prefix: str = "c3a_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: C3AConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/fourierft/model.py

@@ -58,9 +58,6 @@ class FourierFTModel(BaseTuner):
 
     prefix: str = "fourierft_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: FourierFTConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/ia3/model.py

@@ -75,9 +75,6 @@ class IA3Model(BaseTuner):
 
     prefix: str = "ia3_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False):
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     @staticmethod
     def _create_new_module(ia3_config, adapter_name, target, **kwargs):
         # avoid eager bnb import

src/peft/tuners/ln_tuning/model.py

@@ -65,10 +65,6 @@ class LNTuningModel(BaseTuner):
 
     prefix: str = "ln_tuning_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        # self.adapter_name = adapter_name
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def __getattr__(self, name: str):
         """Forward missing attributes to the wrapped module."""
         try:

src/peft/tuners/loha/model.py

@@ -18,6 +18,7 @@
 from torch import nn
 
 from peft.tuners.lycoris_utils import LycorisConfig, LycorisTuner
+from peft.utils import TRANSFORMERS_MODELS_TO_LOHA_TARGET_MODULES_MAPPING
 from peft.utils.other import get_pattern_key
 
 from .layer import Conv2d, Linear, LoHaLayer
@@ -110,3 +111,13 @@ def _create_and_replace(
         else:
             new_module = self._create_new_module(config, adapter_name, target, **kwargs)
             self._replace_module(parent, target_name, new_module, target)
+
+    @staticmethod
+    def _prepare_adapter_config(peft_config, model_config):
+        if peft_config.target_modules is None:
+            if model_config["model_type"] not in TRANSFORMERS_MODELS_TO_LOHA_TARGET_MODULES_MAPPING:
+                raise ValueError("Please specify `target_modules` in `peft_config`")
+            peft_config.target_modules = set(
+                TRANSFORMERS_MODELS_TO_LOHA_TARGET_MODULES_MAPPING[model_config["model_type"]]
+            )
+        return peft_config

src/peft/tuners/lokr/model.py

@@ -18,6 +18,7 @@
 from torch import nn
 
 from peft.tuners.lycoris_utils import LycorisConfig, LycorisTuner
+from peft.utils import TRANSFORMERS_MODELS_TO_LOKR_TARGET_MODULES_MAPPING
 from peft.utils.other import get_pattern_key
 
 from .layer import Conv2d, Linear, LoKrLayer
@@ -112,3 +113,13 @@ def _create_and_replace(
         else:
             new_module = self._create_new_module(config, adapter_name, target, **kwargs)
             self._replace_module(parent, target_name, new_module, target)
+
+    @staticmethod
+    def _prepare_adapter_config(peft_config, model_config):
+        if peft_config.target_modules is None:
+            if model_config["model_type"] not in TRANSFORMERS_MODELS_TO_LOKR_TARGET_MODULES_MAPPING:
+                raise ValueError("Please specify `target_modules` in `peft_config`")
+            peft_config.target_modules = set(
+                TRANSFORMERS_MODELS_TO_LOKR_TARGET_MODULES_MAPPING[model_config["model_type"]]
+            )
+        return peft_config

src/peft/tuners/lora/model.py

@@ -139,9 +139,6 @@ class LoraModel(BaseTuner):
 
     prefix: str = "lora_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: LoraConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/lycoris_utils.py

@@ -203,9 +203,6 @@ class LycorisTuner(BaseTuner):
     prefix: str
     layers_mapping: dict[type[torch.nn.Module], type[LycorisLayer]]
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False):
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def __getattr__(self, name: str):
         """Forward missing attributes to the wrapped module."""
         try:

src/peft/tuners/oft/model.py

@@ -99,9 +99,6 @@ class OFTModel(BaseTuner):
 
     prefix: str = "oft_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: OFTConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/poly/model.py

@@ -33,9 +33,6 @@
 class PolyModel(BaseTuner):
     prefix: str = "poly_"
 
-    def __init__(self, model, config, adapter_name) -> None:
-        super().__init__(model, config, adapter_name)
-
     @staticmethod
     def _check_target_module_exists(poly_config, key):
         return check_target_module_exists(poly_config, key)

src/peft/tuners/randlora/model.py

@@ -101,9 +101,6 @@ class RandLoraModel(BaseTuner):
 
     prefix: str = "randlora_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _find_dim(self, config) -> tuple[int, int]:
         """
         Finds the largest input and output dimensions across linear layers that have been wrapped with RandLora.

src/peft/tuners/shira/model.py

@@ -64,9 +64,6 @@ class ShiraModel(BaseTuner):
 
     prefix: str = "shira_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False) -> None:
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def _check_new_adapter_config(self, config: ShiraConfig) -> None:
         """
         A helper method to check the config when a new adapter is being added.

src/peft/tuners/trainable_tokens/model.py

@@ -31,9 +31,6 @@
 class TrainableTokensModel(BaseTuner):
     prefix: str = "trainable_tokens_"
 
-    def __init__(self, model, config, adapter_name, low_cpu_mem_usage: bool = False):
-        super().__init__(model, config, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage)
-
     def __getattr__(self, name: str):
         """Forward missing attributes to the wrapped module."""
         try:
@@ -49,13 +46,19 @@ def _prepare_adapter_config(self, peft_config, model_config):
         return peft_config
 
     def inject_adapter(
-        self, model: nn.Module, adapter_name: str, autocast_adapter_dtype: bool = True, low_cpu_mem_usage: bool = False
+        self,
+        model: nn.Module,
+        adapter_name: str,
+        autocast_adapter_dtype: bool = True,
+        low_cpu_mem_usage: bool = False,
+        **kwargs,
     ) -> None:
         super().inject_adapter(
             model=model,
             adapter_name=adapter_name,
             autocast_adapter_dtype=autocast_adapter_dtype,
             low_cpu_mem_usage=low_cpu_mem_usage,
+            **kwargs,
         )
 
         model_config = self.get_model_config(self)