peft/tuners/boft/model.py

# Copyright 2023-present the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# The implementation is based on "Parameter-Efficient Orthogonal Finetuning
# via Butterfly Factorization" (https://arxiv.org/abs/2311.06243) in ICLR 2024.

import warnings
from dataclasses import asdict
from enum import Enum
from typing import List, Optional

import torch
from torch import nn
from tqdm import tqdm

from peft.tuners.tuners_utils import BaseTuner, BaseTunerLayer, check_target_module_exists
from peft.utils import (
    TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING,
    ModulesToSaveWrapper,
    _get_submodules,
)

from .config import BOFTConfig
from .layer import BOFTLayer, Conv2d, Linear


class BOFTModel(BaseTuner):
    """
    Creates BOFT and OFT model from a pretrained transformers model. Paper: https://arxiv.org/abs/2311.06243
    https://arxiv.org/abs/2306.07280

    Args:
        model ([`transformers.PreTrainedModel`]): The model to be adapted.
        config ([`BOFTConfig`]): The configuration of the BOFT model.
        adapter_name (`str`): The name of the adapter, defaults to `"default"`.

    Returns:
        `torch.nn.Module`: The BOFT model.

    Example::

        >>> import transformers >>> from transformers import AutoModelForSeq2SeqLM, BOFTConfig >>> from peft import
        BOFTConfig, get_peft_model

        >>> config = BOFTConfig( ... boft_block_size=8, ... boft_n_butterfly_factor=1, ... target_modules=["query",
        "value", "key", "output.dense", "mlp.fc1", "mlp.fc2"], ... boft_dropout=0.1, ... bias="boft_only", ...
        modules_to_save=["classifier"], ... )

        >>> model = transformers.Dinov2ForImageClassification.from_pretrained( ... "facebook/dinov2-large", ...
        num_labels=100, ... ) >>> boft_model = get_peft_model(model, config)

    **Attributes**:
        - **model** ([`transformers.PreTrainedModel`]) -- The model to be adapted.
        - **peft_config** ([`BOFTConfig`]): The configuration of the BOFT model.
    """

    prefix: str = "boft_"

    def __init__(self, model, config, adapter_name) -> None:
        super().__init__(model, config, adapter_name)

    def _check_new_adapter_config(self, config: BOFTConfig) -> None:
        """
        A helper method to check the config when a new adapter is being added.

        Raise a ValueError if there is something wrong with the config or if it conflicts with existing adapters.

        """
        # TODO: there should be a check if any of the existing adapters actually has bias != "none", or else the check
        # does not fully correspond to the error message.
        if (len(self.peft_config) > 1) and (config.bias != "none"):
            raise ValueError(
                f"{self.__class__.__name__} supports only 1 adapter with bias. When using multiple adapters, "
                "set bias to 'none' for all adapters."
            )

    @staticmethod
    def _check_target_module_exists(boft_config, key):
        return check_target_module_exists(boft_config, key)

    def _create_and_replace(
        self,
        boft_config,
        adapter_name,
        target,
        target_name,
        parent,
        current_key,
        **optional_kwargs,
    ):
        if current_key is None:
            raise ValueError("Current Key shouldn't be `None`")

        bias = hasattr(target, "bias") and target.bias is not None
        kwargs = {
            "boft_block_size": boft_config.boft_block_size,
            "boft_block_num": boft_config.boft_block_num,
            "boft_n_butterfly_factor": boft_config.boft_n_butterfly_factor,
            "boft_dropout": boft_config.boft_dropout,
            "fan_in_fan_out": boft_config.fan_in_fan_out,
            "init_weights": boft_config.init_weights,
        }
        kwargs["bias"] = bias

        # If it is not a BOFTLayer, create a new module, else update it with new adapters
        if not isinstance(target, BOFTLayer):
            new_module = self._create_new_module(boft_config, adapter_name, target, **kwargs)
            if adapter_name not in self.active_adapters:
                # adding an additional adapter: it is not automatically trainable
                new_module.requires_grad_(False)
            self._replace_module(parent, target_name, new_module, target)
        else:
            target.update_layer(
                adapter_name,
                boft_block_size=boft_config.boft_block_size,
                boft_block_num=boft_config.boft_block_num,
                boft_n_butterfly_factor=boft_config.boft_n_butterfly_factor,
                boft_dropout=boft_config.boft_dropout,
                init_weights=boft_config.init_weights,
            )

    def _replace_module(self, parent, child_name, new_module, child):
        setattr(parent, child_name, new_module)
        # It's not necessary to set requires_grad here, as that is handled by
        # _mark_only_adapters_as_trainable

        # child layer wraps the original module, unpack it
        if hasattr(child, "base_layer"):
            child = child.base_layer

        if not hasattr(new_module, "base_layer"):
            new_module.weight = child.weight
            if hasattr(child, "bias"):
                new_module.bias = child.bias

        if getattr(child, "state", None) is not None:
            if hasattr(new_module, "base_layer"):
                new_module.base_layer.state = child.state
            else:
                new_module.state = child.state
            new_module.to(child.weight.device)

        # dispatch to correct device
        for name, module in new_module.named_modules():
            if self.prefix in name:
                module.to(child.weight.device)

    def _mark_only_adapters_as_trainable(self, model: nn.Module) -> None:
        for n, p in model.named_parameters():
            if self.prefix not in n:
                p.requires_grad = False

        for active_adapter in self.active_adapters:
            bias = self.peft_config[active_adapter].bias
            if bias == "none":
                continue

            if bias == "all":
                for n, p in model.named_parameters():
                    if "bias" in n:
                        p.requires_grad = True
            elif bias == "boft_only":
                for name, m in model.named_modules():
                    if isinstance(m, BOFTLayer) and hasattr(m, "bias") and m.bias is not None:
                        m.bias.requires_grad = True
            else:
                raise NotImplementedError(f"Requested bias: {bias}, is not implemented.")

    @staticmethod
    def _create_new_module(boft_config, adapter_name, target, **kwargs):
        if isinstance(target, BaseTunerLayer):
            target_base_layer = target.get_base_layer()
        else:
            target_base_layer = target

        if isinstance(target_base_layer, torch.nn.Linear):
            if kwargs["fan_in_fan_out"]:
                warnings.warn(
                    "fan_in_fan_out is set to True but the target module is `torch.nn.Linear`. "
                    "Setting fan_in_fan_out to False."
                )
                kwargs["fan_in_fan_out"] = boft_config.fan_in_fan_out = False
            new_module = Linear(target, adapter_name, **kwargs)
        elif isinstance(target_base_layer, torch.nn.Conv2d):
            new_module = Conv2d(target, adapter_name, **kwargs)
        else:
            raise ValueError(
                f"Target module {target} is not supported. "
                "Currently, only `torch.nn.Linear` and `torch.nn.Conv2d` are supported."
            )

        return new_module

    def __getattr__(self, name: str):
        """Forward missing attributes to the wrapped module."""
        try:
            return super().__getattr__(name)  # defer to nn.Module's logic
        except AttributeError:
            return getattr(self.model, name)

    def get_peft_config_as_dict(self, inference: bool = False):
        config_dict = {}
        for key, value in self.peft_config.items():
            config = {k: v.value if isinstance(v, Enum) else v for k, v in asdict(value).items()}
            if inference:
                config["inference_mode"] = True
        config_dict[key] = config
        return config

    def _set_adapter_layers(self, enabled=True):
        for module in self.model.modules():
            if isinstance(module, (BaseTunerLayer, ModulesToSaveWrapper)):
                module.enable_adapters(enabled)

    def enable_adapter_layers(self):
        self._set_adapter_layers(enabled=True)

    def disable_adapter_layers(self):
        for active_adapter in self.active_adapters:
            val = self.peft_config[active_adapter].bias
            if val != "none":
                msg = (
                    f"Careful, disabling adapter layers with bias configured to be '{val}' does not produce the same "
                    "output as the the base model would without adaption."
                )
                warnings.warn(msg)
        self._set_adapter_layers(enabled=False)

    def set_adapter(self, adapter_name):
        for module in self.model.modules():
            if isinstance(module, BOFTLayer):
                if module.merged:
                    warnings.warn("Adapter cannot be set when the model is merged. Unmerging the model first.")
                    module.unmerge()
                module.set_adapter(adapter_name)
        self.active_adapter = adapter_name

    @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_LORA_TARGET_MODULES_MAPPING:
                raise ValueError("Please specify `target_modules` in `peft_config`")
            peft_config.target_modules = set(
                TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING[model_config["model_type"]]
            )
        return peft_config

    def _unload_and_optionally_merge(
        self,
        merge=True,
        progressbar: bool = False,
        safe_merge: bool = False,
        adapter_names: Optional[List[str]] = None,
    ):
        self._unloading_checks(adapter_names)
        key_list = [key for key, _ in self.model.named_modules() if self.prefix not in key]
        desc = "Unloading " + ("and merging " if merge else "") + "model"
        for key in tqdm(key_list, disable=not progressbar, desc=desc):
            try:
                parent, target, target_name = _get_submodules(self.model, key)
            except AttributeError:
                continue

            if hasattr(target, "base_layer"):
                if merge:
                    target.merge(safe_merge=safe_merge, adapter_names=adapter_names)
                self._replace_module(parent, target_name, target.get_base_layer(), target)
            elif isinstance(target, ModulesToSaveWrapper):
                # save any additional trainable modules part of `modules_to_save`
                setattr(parent, target_name, target.modules_to_save[target.active_adapter])

        return self.model

    def delete_adapter(self, adapter_name: str) -> None:
        """
        Deletes an existing adapter.

        Args:
            adapter_name (str): Name of the adapter to be deleted.
        """
        if adapter_name not in list(self.peft_config.keys()):
            raise ValueError(f"Adapter {adapter_name} does not exist")
        del self.peft_config[adapter_name]

        key_list = [key for key, _ in self.model.named_modules() if self.prefix not in key]
        new_adapter = None
        for key in key_list:
            _, target, _ = _get_submodules(self.model, key)
            if isinstance(target, BOFTLayer):
                target.delete_adapter(adapter_name)
                if new_adapter is None:
                    new_adapter = target.active_adapters[:]

        self.active_adapter = new_adapter or []

    def merge_and_unload(
        self, progressbar: bool = False, safe_merge: bool = False, adapter_names: Optional[List[str]] = None
    ) -> torch.nn.Module:
        r"""
        This method merges the BOFT layers into the base model. This is needed if someone wants to use the base model
        as a standalone model.

        Args:
            progressbar (`bool`):
                whether to show a progressbar indicating the unload and merge process
            safe_merge (`bool`):
                whether to activate the safe merging check to check if there is any potential Nan in the adapter
                weights
            adapter_names (`List[str]`, *optional*):
                The list of adapter names that should be merged. If None, all active adapters will be merged. Defaults
                to `None`.

        """
        return self._unload_and_optionally_merge(
            progressbar=progressbar, safe_merge=safe_merge, adapter_names=adapter_names
        )

    def unload(self) -> torch.nn.Module:
        """
        Gets back the base model by removing all the boft modules without merging. This gives back the original base
        model.
        """
        return self._unload_and_optionally_merge(merge=False)