# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# 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.
import copy
import inspect
import os
from pathlib import Path
from typing import Callable, Dict, List, Optional, Union
import safetensors
import torch
from huggingface_hub import model_info
from huggingface_hub.constants import HF_HUB_OFFLINE
from huggingface_hub.utils import validate_hf_hub_args
from torch import nn
from ..models.modeling_utils import load_state_dict
from ..utils import (
USE_PEFT_BACKEND,
_get_model_file,
convert_state_dict_to_diffusers,
convert_state_dict_to_peft,
convert_unet_state_dict_to_peft,
delete_adapter_layers,
get_adapter_name,
get_peft_kwargs,
is_accelerate_available,
is_peft_version,
is_transformers_available,
logging,
recurse_remove_peft_layers,
scale_lora_layers,
set_adapter_layers,
set_weights_and_activate_adapters,
)
from .lora_conversion_utils import _convert_non_diffusers_lora_to_diffusers, _maybe_map_sgm_blocks_to_diffusers
if is_transformers_available():
from transformers import PreTrainedModel
from ..models.lora import text_encoder_attn_modules, text_encoder_mlp_modules
if is_accelerate_available():
from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
logger = logging.get_logger(__name__)
TEXT_ENCODER_NAME = "text_encoder"
UNET_NAME = "unet"
TRANSFORMER_NAME = "transformer"
LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
class LoraLoaderMixin:
r"""
Load LoRA layers into [`UNet2DConditionModel`] and
[`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
"""
text_encoder_name = TEXT_ENCODER_NAME
unet_name = UNET_NAME
transformer_name = TRANSFORMER_NAME
num_fused_loras = 0
def load_lora_weights(
self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
):
"""
Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
`self.text_encoder`.
All kwargs are forwarded to `self.lora_state_dict`.
See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
`self.unet`.
See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
into `self.text_encoder`.
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
kwargs (`dict`, *optional*):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# if a dict is passed, copy it instead of modifying it inplace
if isinstance(pretrained_model_name_or_path_or_dict, dict):
pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")
self.load_lora_into_unet(
state_dict,
network_alphas=network_alphas,
unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
adapter_name=adapter_name,
_pipeline=self,
)
self.load_lora_into_text_encoder(
state_dict,
network_alphas=network_alphas,
text_encoder=getattr(self, self.text_encoder_name)
if not hasattr(self, "text_encoder")
else self.text_encoder,
lora_scale=self.lora_scale,
adapter_name=adapter_name,
_pipeline=self,
)
@classmethod
@validate_hf_hub_args
def lora_state_dict(
cls,
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
**kwargs,
):
r"""
Return state dict for lora weights and the network alphas.
<Tip warning={true}>
We support loading A1111 formatted LoRA checkpoints in a limited capacity.
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
the Hub.
- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
with [`ModelMixin.save_pretrained`].
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
is not used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download:
Deprecated and ignored. All downloads are now resumed by default when possible. Will be removed in v1
of Diffusers.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only (`bool`, *optional*, defaults to `False`):
Whether to only load local model weights and configuration files or not. If set to `True`, the model
won't be downloaded from the Hub.
token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
`diffusers-cli login` (stored in `~/.huggingface`) is used.
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
allowed by Git.
subfolder (`str`, *optional*, defaults to `""`):
The subfolder location of a model file within a larger model repository on the Hub or locally.
weight_name (`str`, *optional*, defaults to None):
Name of the serialized state dict file.
"""
# Load the main state dict first which has the LoRA layers for either of
# UNet and text encoder or both.
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", None)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", None)
token = kwargs.pop("token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
unet_config = kwargs.pop("unet_config", None)
use_safetensors = kwargs.pop("use_safetensors", None)
allow_pickle = False
if use_safetensors is None:
use_safetensors = True
allow_pickle = True
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
model_file = None
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
# Here we're relaxing the loading check to enable more Inference API
# friendliness where sometimes, it's not at all possible to automatically
# determine `weight_name`.
if weight_name is None:
weight_name = cls._best_guess_weight_name(
pretrained_model_name_or_path_or_dict,
file_extension=".safetensors",
local_files_only=local_files_only,
)
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except (IOError, safetensors.SafetensorError) as e:
if not allow_pickle:
raise e
# try loading non-safetensors weights
model_file = None
pass
if model_file is None:
if weight_name is None:
weight_name = cls._best_guess_weight_name(
pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
)
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = load_state_dict(model_file)
else:
state_dict = pretrained_model_name_or_path_or_dict
network_alphas = None
# TODO: replace it with a method from `state_dict_utils`
if all(
(
k.startswith("lora_te_")
or k.startswith("lora_unet_")
or k.startswith("lora_te1_")
or k.startswith("lora_te2_")
)
for k in state_dict.keys()
):
# Map SDXL blocks correctly.
if unet_config is not None:
# use unet config to remap block numbers
state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
state_dict, network_alphas = _convert_non_diffusers_lora_to_diffusers(state_dict)
return state_dict, network_alphas
@classmethod
def _best_guess_weight_name(
cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
):
if local_files_only or HF_HUB_OFFLINE:
raise ValueError("When using the offline mode, you must specify a `weight_name`.")
targeted_files = []
if os.path.isfile(pretrained_model_name_or_path_or_dict):
return
elif os.path.isdir(pretrained_model_name_or_path_or_dict):
targeted_files = [
f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
]
else:
files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
if len(targeted_files) == 0:
return
# "scheduler" does not correspond to a LoRA checkpoint.
# "optimizer" does not correspond to a LoRA checkpoint
# only top-level checkpoints are considered and not the other ones, hence "checkpoint".
unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
targeted_files = list(
filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
)
if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
if len(targeted_files) > 1:
raise ValueError(
f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one `.safetensors` or `.bin` file in {pretrained_model_name_or_path_or_dict}."
)
weight_name = targeted_files[0]
return weight_name
@classmethod
def _optionally_disable_offloading(cls, _pipeline):
"""
Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
Args:
_pipeline (`DiffusionPipeline`):
The pipeline to disable offloading for.
Returns:
tuple:
A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
"""
is_model_cpu_offload = False
is_sequential_cpu_offload = False
if _pipeline is not None and _pipeline.hf_device_map is None:
for _, component in _pipeline.components.items():
if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
if not is_model_cpu_offload:
is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
if not is_sequential_cpu_offload:
is_sequential_cpu_offload = (
isinstance(component._hf_hook, AlignDevicesHook)
or hasattr(component._hf_hook, "hooks")
and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
)
logger.info(
"Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
)
remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
return (is_model_cpu_offload, is_sequential_cpu_offload)
@classmethod
def load_lora_into_unet(cls, state_dict, network_alphas, unet, adapter_name=None, _pipeline=None):
"""
This will load the LoRA layers specified in `state_dict` into `unet`.
Parameters:
state_dict (`dict`):
A standard state dict containing the lora layer parameters. The keys can either be indexed directly
into the unet or prefixed with an additional `unet` which can be used to distinguish between text
encoder lora layers.
network_alphas (`Dict[str, float]`):
The value of the network alpha used for stable learning and preventing underflow. This value has the
same meaning as the `--network_alpha` option in the kohya-ss trainer script. Refer to [this
link](https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning).
unet (`UNet2DConditionModel`):
The UNet model to load the LoRA layers into.
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
# then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
# their prefixes.
keys = list(state_dict.keys())
only_text_encoder = all(key.startswith(cls.text_encoder_name) for key in keys)
if any(key.startswith(cls.unet_name) for key in keys) and not only_text_encoder:
# Load the layers corresponding to UNet.
logger.info(f"Loading {cls.unet_name}.")
unet.load_attn_procs(
state_dict, network_alphas=network_alphas, adapter_name=adapter_name, _pipeline=_pipeline
)
@classmethod
def load_lora_into_text_encoder(
cls,
state_dict,
network_alphas,
text_encoder,
prefix=None,
lora_scale=1.0,
adapter_name=None,
_pipeline=None,
):
"""
This will load the LoRA layers specified in `state_dict` into `text_encoder`
Parameters:
state_dict (`dict`):
A standard state dict containing the lora layer parameters. The key should be prefixed with an
additional `text_encoder` to distinguish between unet lora layers.
network_alphas (`Dict[str, float]`):
See `LoRALinearLayer` for more details.
text_encoder (`CLIPTextModel`):
The text encoder model to load the LoRA layers into.
prefix (`str`):
Expected prefix of the `text_encoder` in the `state_dict`.
lora_scale (`float`):
How much to scale the output of the lora linear layer before it is added with the output of the regular
lora layer.
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
from peft import LoraConfig
# If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
# then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
# their prefixes.
keys = list(state_dict.keys())
prefix = cls.text_encoder_name if prefix is None else prefix
# Safe prefix to check with.
if any(cls.text_encoder_name in key for key in keys):
# Load the layers corresponding to text encoder and make necessary adjustments.
text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
text_encoder_lora_state_dict = {
k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
}
if len(text_encoder_lora_state_dict) > 0:
logger.info(f"Loading {prefix}.")
rank = {}
text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)
# convert state dict
text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
for name, _ in text_encoder_attn_modules(text_encoder):
for module in ("out_proj", "q_proj", "k_proj", "v_proj"):
rank_key = f"{name}.{module}.lora_B.weight"
if rank_key not in text_encoder_lora_state_dict:
continue
rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
for name, _ in text_encoder_mlp_modules(text_encoder):
for module in ("fc1", "fc2"):
rank_key = f"{name}.{module}.lora_B.weight"
if rank_key not in text_encoder_lora_state_dict:
continue
rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
if network_alphas is not None:
alpha_keys = [
k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
]
network_alphas = {
k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
}
lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
if "use_dora" in lora_config_kwargs:
if lora_config_kwargs["use_dora"]:
if is_peft_version("<", "0.9.0"):
raise ValueError(
"You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
)
else:
if is_peft_version("<", "0.9.0"):
lora_config_kwargs.pop("use_dora")
lora_config = LoraConfig(**lora_config_kwargs)
# adapter_name
if adapter_name is None:
adapter_name = get_adapter_name(text_encoder)
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
# inject LoRA layers and load the state dict
# in transformers we automatically check whether the adapter name is already in use or not
text_encoder.load_adapter(
adapter_name=adapter_name,
adapter_state_dict=text_encoder_lora_state_dict,
peft_config=lora_config,
)
# scale LoRA layers with `lora_scale`
scale_lora_layers(text_encoder, weight=lora_scale)
text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
# Unsafe code />
@classmethod
def load_lora_into_transformer(cls, state_dict, network_alphas, transformer, adapter_name=None, _pipeline=None):
"""
This will load the LoRA layers specified in `state_dict` into `transformer`.
Parameters:
state_dict (`dict`):
A standard state dict containing the lora layer parameters. The keys can either be indexed directly
into the unet or prefixed with an additional `unet` which can be used to distinguish between text
encoder lora layers.
network_alphas (`Dict[str, float]`):
See `LoRALinearLayer` for more details.
unet (`UNet2DConditionModel`):
The UNet model to load the LoRA layers into.
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
keys = list(state_dict.keys())
transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
state_dict = {
k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
}
if network_alphas is not None:
alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.transformer_name)]
network_alphas = {
k.replace(f"{cls.transformer_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
}
if len(state_dict.keys()) > 0:
if adapter_name in getattr(transformer, "peft_config", {}):
raise ValueError(
f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
)
rank = {}
for key, val in state_dict.items():
if "lora_B" in key:
rank[key] = val.shape[1]
lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict)
if "use_dora" in lora_config_kwargs:
if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
raise ValueError(
"You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
)
else:
lora_config_kwargs.pop("use_dora")
lora_config = LoraConfig(**lora_config_kwargs)
# adapter_name
if adapter_name is None:
adapter_name = get_adapter_name(transformer)
# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
logger.warning(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
# Unsafe code />
@property
def lora_scale(self) -> float:
# property function that returns the lora scale which can be set at run time by the pipeline.
# if _lora_scale has not been set, return 1
return self._lora_scale if hasattr(self, "_lora_scale") else 1.0
def _remove_text_encoder_monkey_patch(self):
remove_method = recurse_remove_peft_layers
if hasattr(self, "text_encoder"):
remove_method(self.text_encoder)
# In case text encoder have no Lora attached
if getattr(self.text_encoder, "peft_config", None) is not None:
del self.text_encoder.peft_config
self.text_encoder._hf_peft_config_loaded = None
if hasattr(self, "text_encoder_2"):
remove_method(self.text_encoder_2)
if getattr(self.text_encoder_2, "peft_config", None) is not None:
del self.text_encoder_2.peft_config
self.text_encoder_2._hf_peft_config_loaded = None
@classmethod
def save_lora_weights(
cls,
save_directory: Union[str, os.PathLike],
unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
transformer_lora_layers: Dict[str, torch.nn.Module] = None,
is_main_process: bool = True,
weight_name: str = None,
save_function: Callable = None,
safe_serialization: bool = True,
):
r"""
Save the LoRA parameters corresponding to the UNet and text encoder.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to save LoRA parameters to. Will be created if it doesn't exist.
unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `unet`.
text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
encoder LoRA state dict because it comes from 🤗 Transformers.
is_main_process (`bool`, *optional*, defaults to `True`):
Whether the process calling this is the main process or not. Useful during distributed training and you
need to call this function on all processes. In this case, set `is_main_process=True` only on the main
process to avoid race conditions.
save_function (`Callable`):
The function to use to save the state dictionary. Useful during distributed training when you need to
replace `torch.save` with another method. Can be configured with the environment variable
`DIFFUSERS_SAVE_MODE`.
safe_serialization (`bool`, *optional*, defaults to `True`):
Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
"""
state_dict = {}
def pack_weights(layers, prefix):
layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
return layers_state_dict
if not (unet_lora_layers or text_encoder_lora_layers or transformer_lora_layers):
raise ValueError(
"You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers`, or `transformer_lora_layers`."
)
if unet_lora_layers:
state_dict.update(pack_weights(unet_lora_layers, cls.unet_name))
if text_encoder_lora_layers:
state_dict.update(pack_weights(text_encoder_lora_layers, cls.text_encoder_name))
if transformer_lora_layers:
state_dict.update(pack_weights(transformer_lora_layers, "transformer"))
# Save the model
cls.write_lora_layers(
state_dict=state_dict,
save_directory=save_directory,
is_main_process=is_main_process,
weight_name=weight_name,
save_function=save_function,
safe_serialization=safe_serialization,
)
@staticmethod
def write_lora_layers(
state_dict: Dict[str, torch.Tensor],
save_directory: str,
is_main_process: bool,
weight_name: str,
save_function: Callable,
safe_serialization: bool,
):
if os.path.isfile(save_directory):
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
return
if save_function is None:
if safe_serialization:
def save_function(weights, filename):
return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
else:
save_function = torch.save
os.makedirs(save_directory, exist_ok=True)
if weight_name is None:
if safe_serialization:
weight_name = LORA_WEIGHT_NAME_SAFE
else:
weight_name = LORA_WEIGHT_NAME
save_path = Path(save_directory, weight_name).as_posix()
save_function(state_dict, save_path)
logger.info(f"Model weights saved in {save_path}")
def unload_lora_weights(self):
"""
Unloads the LoRA parameters.
Examples:
```python
>>> # Assuming `pipeline` is already loaded with the LoRA parameters.
>>> pipeline.unload_lora_weights()
>>> ...
```
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
unet.unload_lora()
# Safe to call the following regardless of LoRA.
self._remove_text_encoder_monkey_patch()
def fuse_lora(
self,
fuse_unet: bool = True,
fuse_text_encoder: bool = True,
lora_scale: float = 1.0,
safe_fusing: bool = False,
adapter_names: Optional[List[str]] = None,
):
r"""
Fuses the LoRA parameters into the original parameters of the corresponding blocks.
<Tip warning={true}>
This is an experimental API.
</Tip>
Args:
fuse_unet (`bool`, defaults to `True`): Whether to fuse the UNet LoRA parameters.
fuse_text_encoder (`bool`, defaults to `True`):
Whether to fuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
LoRA parameters then it won't have any effect.
lora_scale (`float`, defaults to 1.0):
Controls how much to influence the outputs with the LoRA parameters.
safe_fusing (`bool`, defaults to `False`):
Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
adapter_names (`List[str]`, *optional*):
Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
Example:
```py
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
).to("cuda")
pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
pipeline.fuse_lora(lora_scale=0.7)
```
"""
from peft.tuners.tuners_utils import BaseTunerLayer
if fuse_unet or fuse_text_encoder:
self.num_fused_loras += 1
if self.num_fused_loras > 1:
logger.warning(
"The current API is supported for operating with a single LoRA file. You are trying to load and fuse more than one LoRA which is not well-supported.",
)
if fuse_unet:
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
unet.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, adapter_names=None):
merge_kwargs = {"safe_merge": safe_fusing}
for module in text_encoder.modules():
if isinstance(module, BaseTunerLayer):
if lora_scale != 1.0:
module.scale_layer(lora_scale)
# For BC with previous PEFT versions, we need to check the signature
# of the `merge` method to see if it supports the `adapter_names` argument.
supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
if "adapter_names" in supported_merge_kwargs:
merge_kwargs["adapter_names"] = adapter_names
elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
raise ValueError(
"The `adapter_names` argument is not supported with your PEFT version. "
"Please upgrade to the latest version of PEFT. `pip install -U peft`"
)
module.merge(**merge_kwargs)
if fuse_text_encoder:
if hasattr(self, "text_encoder"):
fuse_text_encoder_lora(self.text_encoder, lora_scale, safe_fusing, adapter_names=adapter_names)
if hasattr(self, "text_encoder_2"):
fuse_text_encoder_lora(self.text_encoder_2, lora_scale, safe_fusing, adapter_names=adapter_names)
def unfuse_lora(self, unfuse_unet: bool = True, unfuse_text_encoder: bool = True):
r"""
Reverses the effect of
[`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
<Tip warning={true}>
This is an experimental API.
</Tip>
Args:
unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
unfuse_text_encoder (`bool`, defaults to `True`):
Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
LoRA parameters then it won't have any effect.
"""
from peft.tuners.tuners_utils import BaseTunerLayer
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
if unfuse_unet:
for module in unet.modules():
if isinstance(module, BaseTunerLayer):
module.unmerge()
def unfuse_text_encoder_lora(text_encoder):
for module in text_encoder.modules():
if isinstance(module, BaseTunerLayer):
module.unmerge()
if unfuse_text_encoder:
if hasattr(self, "text_encoder"):
unfuse_text_encoder_lora(self.text_encoder)
if hasattr(self, "text_encoder_2"):
unfuse_text_encoder_lora(self.text_encoder_2)
self.num_fused_loras -= 1
def set_adapters_for_text_encoder(
self,
adapter_names: Union[List[str], str],
text_encoder: Optional["PreTrainedModel"] = None, # noqa: F821
text_encoder_weights: Optional[Union[float, List[float], List[None]]] = None,
):
"""
Sets the adapter layers for the text encoder.
Args:
adapter_names (`List[str]` or `str`):
The names of the adapters to use.
text_encoder (`torch.nn.Module`, *optional*):
The text encoder module to set the adapter layers for. If `None`, it will try to get the `text_encoder`
attribute.
text_encoder_weights (`List[float]`, *optional*):
The weights to use for the text encoder. If `None`, the weights are set to `1.0` for all the adapters.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
def process_weights(adapter_names, weights):
# Expand weights into a list, one entry per adapter
# e.g. for 2 adapters: 7 -> [7,7] ; [3, None] -> [3, None]
if not isinstance(weights, list):
weights = [weights] * len(adapter_names)
if len(adapter_names) != len(weights):
raise ValueError(
f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
)
# Set None values to default of 1.0
# e.g. [7,7] -> [7,7] ; [3, None] -> [3,1]
weights = [w if w is not None else 1.0 for w in weights]
return weights
adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
text_encoder_weights = process_weights(adapter_names, text_encoder_weights)
text_encoder = text_encoder or getattr(self, "text_encoder", None)
if text_encoder is None:
raise ValueError(
"The pipeline does not have a default `pipe.text_encoder` class. Please make sure to pass a `text_encoder` instead."
)
set_weights_and_activate_adapters(text_encoder, adapter_names, text_encoder_weights)
def disable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
"""
Disables the LoRA layers for the text encoder.
Args:
text_encoder (`torch.nn.Module`, *optional*):
The text encoder module to disable the LoRA layers for. If `None`, it will try to get the
`text_encoder` attribute.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
text_encoder = text_encoder or getattr(self, "text_encoder", None)
if text_encoder is None:
raise ValueError("Text Encoder not found.")
set_adapter_layers(text_encoder, enabled=False)
def enable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
"""
Enables the LoRA layers for the text encoder.
Args:
text_encoder (`torch.nn.Module`, *optional*):
The text encoder module to enable the LoRA layers for. If `None`, it will try to get the `text_encoder`
attribute.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
text_encoder = text_encoder or getattr(self, "text_encoder", None)
if text_encoder is None:
raise ValueError("Text Encoder not found.")
set_adapter_layers(self.text_encoder, enabled=True)
def set_adapters(
self,
adapter_names: Union[List[str], str],
adapter_weights: Optional[Union[float, Dict, List[float], List[Dict]]] = None,
):
adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
adapter_weights = copy.deepcopy(adapter_weights)
# Expand weights into a list, one entry per adapter
if not isinstance(adapter_weights, list):
adapter_weights = [adapter_weights] * len(adapter_names)
if len(adapter_names) != len(adapter_weights):
raise ValueError(
f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(adapter_weights)}"
)
# Decompose weights into weights for unet, text_encoder and text_encoder_2
unet_lora_weights, text_encoder_lora_weights, text_encoder_2_lora_weights = [], [], []
list_adapters = self.get_list_adapters() # eg {"unet": ["adapter1", "adapter2"], "text_encoder": ["adapter2"]}
all_adapters = {
adapter for adapters in list_adapters.values() for adapter in adapters
} # eg ["adapter1", "adapter2"]
invert_list_adapters = {
adapter: [part for part, adapters in list_adapters.items() if adapter in adapters]
for adapter in all_adapters
} # eg {"adapter1": ["unet"], "adapter2": ["unet", "text_encoder"]}
for adapter_name, weights in zip(adapter_names, adapter_weights):
if isinstance(weights, dict):
unet_lora_weight = weights.pop("unet", None)
text_encoder_lora_weight = weights.pop("text_encoder", None)
text_encoder_2_lora_weight = weights.pop("text_encoder_2", None)
if len(weights) > 0:
raise ValueError(
f"Got invalid key '{weights.keys()}' in lora weight dict for adapter {adapter_name}."
)
if text_encoder_2_lora_weight is not None and not hasattr(self, "text_encoder_2"):
logger.warning(
"Lora weight dict contains text_encoder_2 weights but will be ignored because pipeline does not have text_encoder_2."
)
# warn if adapter doesn't have parts specified by adapter_weights
for part_weight, part_name in zip(
[unet_lora_weight, text_encoder_lora_weight, text_encoder_2_lora_weight],
["unet", "text_encoder", "text_encoder_2"],
):
if part_weight is not None and part_name not in invert_list_adapters[adapter_name]:
logger.warning(
f"Lora weight dict for adapter '{adapter_name}' contains {part_name}, but this will be ignored because {adapter_name} does not contain weights for {part_name}. Valid parts for {adapter_name} are: {invert_list_adapters[adapter_name]}."
)
else:
unet_lora_weight = weights
text_encoder_lora_weight = weights
text_encoder_2_lora_weight = weights
unet_lora_weights.append(unet_lora_weight)
text_encoder_lora_weights.append(text_encoder_lora_weight)
text_encoder_2_lora_weights.append(text_encoder_2_lora_weight)
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
# Handle the UNET
unet.set_adapters(adapter_names, unet_lora_weights)
# Handle the Text Encoder
if hasattr(self, "text_encoder"):
self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, text_encoder_lora_weights)
if hasattr(self, "text_encoder_2"):
self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, text_encoder_2_lora_weights)
def disable_lora(self):
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# Disable unet adapters
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
unet.disable_lora()
# Disable text encoder adapters
if hasattr(self, "text_encoder"):
self.disable_lora_for_text_encoder(self.text_encoder)
if hasattr(self, "text_encoder_2"):
self.disable_lora_for_text_encoder(self.text_encoder_2)
def enable_lora(self):
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# Enable unet adapters
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
unet.enable_lora()
# Enable text encoder adapters
if hasattr(self, "text_encoder"):
self.enable_lora_for_text_encoder(self.text_encoder)
if hasattr(self, "text_encoder_2"):
self.enable_lora_for_text_encoder(self.text_encoder_2)
def delete_adapters(self, adapter_names: Union[List[str], str]):
"""
Args:
Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
adapter_names (`Union[List[str], str]`):
The names of the adapter to delete. Can be a single string or a list of strings
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
if isinstance(adapter_names, str):
adapter_names = [adapter_names]
# Delete unet adapters
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
unet.delete_adapters(adapter_names)
for adapter_name in adapter_names:
# Delete text encoder adapters
if hasattr(self, "text_encoder"):
delete_adapter_layers(self.text_encoder, adapter_name)
if hasattr(self, "text_encoder_2"):
delete_adapter_layers(self.text_encoder_2, adapter_name)
def get_active_adapters(self) -> List[str]:
"""
Gets the list of the current active adapters.
Example:
```python
from diffusers import DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
).to("cuda")
pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
pipeline.get_active_adapters()
```
"""
if not USE_PEFT_BACKEND:
raise ValueError(
"PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
)
from peft.tuners.tuners_utils import BaseTunerLayer
active_adapters = []
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
for module in unet.modules():
if isinstance(module, BaseTunerLayer):
active_adapters = module.active_adapters
break
return active_adapters
def get_list_adapters(self) -> Dict[str, List[str]]:
"""
Gets the current list of all available adapters in the pipeline.
"""
if not USE_PEFT_BACKEND:
raise ValueError(
"PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
)
set_adapters = {}
if hasattr(self, "text_encoder") and hasattr(self.text_encoder, "peft_config"):
set_adapters["text_encoder"] = list(self.text_encoder.peft_config.keys())
if hasattr(self, "text_encoder_2") and hasattr(self.text_encoder_2, "peft_config"):
set_adapters["text_encoder_2"] = list(self.text_encoder_2.peft_config.keys())
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
if hasattr(self, self.unet_name) and hasattr(unet, "peft_config"):
set_adapters[self.unet_name] = list(self.unet.peft_config.keys())
return set_adapters
def set_lora_device(self, adapter_names: List[str], device: Union[torch.device, str, int]) -> None:
"""
Moves the LoRAs listed in `adapter_names` to a target device. Useful for offloading the LoRA to the CPU in case
you want to load multiple adapters and free some GPU memory.
Args:
adapter_names (`List[str]`):
List of adapters to send device to.
device (`Union[torch.device, str, int]`):
Device to send the adapters to. Can be either a torch device, a str or an integer.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
from peft.tuners.tuners_utils import BaseTunerLayer
# Handle the UNET
unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
for unet_module in unet.modules():
if isinstance(unet_module, BaseTunerLayer):
for adapter_name in adapter_names:
unet_module.lora_A[adapter_name].to(device)
unet_module.lora_B[adapter_name].to(device)
# this is a param, not a module, so device placement is not in-place -> re-assign
if hasattr(unet_module, "lora_magnitude_vector") and unet_module.lora_magnitude_vector is not None:
unet_module.lora_magnitude_vector[adapter_name] = unet_module.lora_magnitude_vector[
adapter_name
].to(device)
# Handle the text encoder
modules_to_process = []
if hasattr(self, "text_encoder"):
modules_to_process.append(self.text_encoder)
if hasattr(self, "text_encoder_2"):
modules_to_process.append(self.text_encoder_2)
for text_encoder in modules_to_process:
# loop over submodules
for text_encoder_module in text_encoder.modules():
if isinstance(text_encoder_module, BaseTunerLayer):
for adapter_name in adapter_names:
text_encoder_module.lora_A[adapter_name].to(device)
text_encoder_module.lora_B[adapter_name].to(device)
# this is a param, not a module, so device placement is not in-place -> re-assign
if (
hasattr(text_encoder_module, "lora_magnitude_vector")
and text_encoder_module.lora_magnitude_vector is not None
):
text_encoder_module.lora_magnitude_vector[
adapter_name
] = text_encoder_module.lora_magnitude_vector[adapter_name].to(device)
class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
"""This class overrides `LoraLoaderMixin` with LoRA loading/saving code that's specific to SDXL"""
# Override to properly handle the loading and unloading of the additional text encoder.
def load_lora_weights(
self,
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
adapter_name: Optional[str] = None,
**kwargs,
):
"""
Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
`self.text_encoder`.
All kwargs are forwarded to `self.lora_state_dict`.
See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
`self.unet`.
See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
into `self.text_encoder`.
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
kwargs (`dict`, *optional*):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# We could have accessed the unet config from `lora_state_dict()` too. We pass
# it here explicitly to be able to tell that it's coming from an SDXL
# pipeline.
# if a dict is passed, copy it instead of modifying it inplace
if isinstance(pretrained_model_name_or_path_or_dict, dict):
pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict, network_alphas = self.lora_state_dict(
pretrained_model_name_or_path_or_dict,
unet_config=self.unet.config,
**kwargs,
)
is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")
self.load_lora_into_unet(
state_dict, network_alphas=network_alphas, unet=self.unet, adapter_name=adapter_name, _pipeline=self
)
text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
if len(text_encoder_state_dict) > 0:
self.load_lora_into_text_encoder(
text_encoder_state_dict,
network_alphas=network_alphas,
text_encoder=self.text_encoder,
prefix="text_encoder",
lora_scale=self.lora_scale,
adapter_name=adapter_name,
_pipeline=self,
)
text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
if len(text_encoder_2_state_dict) > 0:
self.load_lora_into_text_encoder(
text_encoder_2_state_dict,
network_alphas=network_alphas,
text_encoder=self.text_encoder_2,
prefix="text_encoder_2",
lora_scale=self.lora_scale,
adapter_name=adapter_name,
_pipeline=self,
)
@classmethod
def save_lora_weights(
cls,
save_directory: Union[str, os.PathLike],
unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
is_main_process: bool = True,
weight_name: str = None,
save_function: Callable = None,
safe_serialization: bool = True,
):
r"""
Save the LoRA parameters corresponding to the UNet and text encoder.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to save LoRA parameters to. Will be created if it doesn't exist.
unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `unet`.
text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
encoder LoRA state dict because it comes from 🤗 Transformers.
text_encoder_2_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `text_encoder_2`. Must explicitly pass the text
encoder LoRA state dict because it comes from 🤗 Transformers.
is_main_process (`bool`, *optional*, defaults to `True`):
Whether the process calling this is the main process or not. Useful during distributed training and you
need to call this function on all processes. In this case, set `is_main_process=True` only on the main
process to avoid race conditions.
save_function (`Callable`):
The function to use to save the state dictionary. Useful during distributed training when you need to
replace `torch.save` with another method. Can be configured with the environment variable
`DIFFUSERS_SAVE_MODE`.
safe_serialization (`bool`, *optional*, defaults to `True`):
Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
"""
state_dict = {}
def pack_weights(layers, prefix):
layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
return layers_state_dict
if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
raise ValueError(
"You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
)
if unet_lora_layers:
state_dict.update(pack_weights(unet_lora_layers, "unet"))
if text_encoder_lora_layers:
state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
if text_encoder_2_lora_layers:
state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
cls.write_lora_layers(
state_dict=state_dict,
save_directory=save_directory,
is_main_process=is_main_process,
weight_name=weight_name,
save_function=save_function,
safe_serialization=safe_serialization,
)
def _remove_text_encoder_monkey_patch(self):
recurse_remove_peft_layers(self.text_encoder)
# TODO: @younesbelkada handle this in transformers side
if getattr(self.text_encoder, "peft_config", None) is not None:
del self.text_encoder.peft_config
self.text_encoder._hf_peft_config_loaded = None
recurse_remove_peft_layers(self.text_encoder_2)
if getattr(self.text_encoder_2, "peft_config", None) is not None:
del self.text_encoder_2.peft_config
self.text_encoder_2._hf_peft_config_loaded = None
class SD3LoraLoaderMixin:
r"""
Load LoRA layers into [`SD3Transformer2DModel`].
"""
transformer_name = TRANSFORMER_NAME
num_fused_loras = 0
def load_lora_weights(
self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
):
"""
Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
`self.text_encoder`.
All kwargs are forwarded to `self.lora_state_dict`.
See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
See [`~loaders.LoraLoaderMixin.load_lora_into_transformer`] for more details on how the state dict is loaded
into `self.transformer`.
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
kwargs (`dict`, *optional*):
See [`~loaders.LoraLoaderMixin.lora_state_dict`].
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
if not USE_PEFT_BACKEND:
raise ValueError("PEFT backend is required for this method.")
# if a dict is passed, copy it instead of modifying it inplace
if isinstance(pretrained_model_name_or_path_or_dict, dict):
pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")
self.load_lora_into_transformer(
state_dict,
transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
adapter_name=adapter_name,
_pipeline=self,
)
@classmethod
@validate_hf_hub_args
def lora_state_dict(
cls,
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
**kwargs,
):
r"""
Return state dict for lora weights and the network alphas.
<Tip warning={true}>
We support loading A1111 formatted LoRA checkpoints in a limited capacity.
This function is experimental and might change in the future.
</Tip>
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
the Hub.
- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
with [`ModelMixin.save_pretrained`].
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
cache_dir (`Union[str, os.PathLike]`, *optional*):
Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
is not used.
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
incompletely downloaded files are deleted.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
local_files_only (`bool`, *optional*, defaults to `False`):
Whether to only load local model weights and configuration files or not. If set to `True`, the model
won't be downloaded from the Hub.
token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
`diffusers-cli login` (stored in `~/.huggingface`) is used.
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
allowed by Git.
subfolder (`str`, *optional*, defaults to `""`):
The subfolder location of a model file within a larger model repository on the Hub or locally.
"""
# Load the main state dict first which has the LoRA layers for either of
# UNet and text encoder or both.
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", None)
token = kwargs.pop("token", None)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", None)
weight_name = kwargs.pop("weight_name", None)
use_safetensors = kwargs.pop("use_safetensors", None)
allow_pickle = False
if use_safetensors is None:
use_safetensors = True
allow_pickle = True
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
model_file = None
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
# Let's first try to load .safetensors weights
if (use_safetensors and weight_name is None) or (
weight_name is not None and weight_name.endswith(".safetensors")
):
try:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = safetensors.torch.load_file(model_file, device="cpu")
except (IOError, safetensors.SafetensorError) as e:
if not allow_pickle:
raise e
# try loading non-safetensors weights
model_file = None
pass
if model_file is None:
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name or LORA_WEIGHT_NAME,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
state_dict = load_state_dict(model_file)
else:
state_dict = pretrained_model_name_or_path_or_dict
return state_dict
@classmethod
def load_lora_into_transformer(cls, state_dict, transformer, adapter_name=None, _pipeline=None):
"""
This will load the LoRA layers specified in `state_dict` into `transformer`.
Parameters:
state_dict (`dict`):
A standard state dict containing the lora layer parameters. The keys can either be indexed directly
into the unet or prefixed with an additional `unet` which can be used to distinguish between text
encoder lora layers.
transformer (`SD3Transformer2DModel`):
The Transformer model to load the LoRA layers into.
adapter_name (`str`, *optional*):
Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
`default_{i}` where i is the total number of adapters being loaded.
"""
from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
keys = list(state_dict.keys())
transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
state_dict = {
k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
}
if len(state_dict.keys()) > 0:
# check with first key if is not in peft format
first_key = next(iter(state_dict.keys()))
if "lora_A" not in first_key:
state_dict = convert_unet_state_dict_to_peft(state_dict)
if adapter_name in getattr(transformer, "peft_config", {}):
raise ValueError(
f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
)
rank = {}
for key, val in state_dict.items():
if "lora_B" in key:
rank[key] = val.shape[1]
lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=None, peft_state_dict=state_dict)
if "use_dora" in lora_config_kwargs:
if lora_config_kwargs["use_dora"] and is_peft_version("<", "0.9.0"):
raise ValueError(
"You need `peft` 0.9.0 at least to use DoRA-enabled LoRAs. Please upgrade your installation of `peft`."
)
else:
lora_config_kwargs.pop("use_dora")
lora_config = LoraConfig(**lora_config_kwargs)
# adapter_name
if adapter_name is None:
adapter_name = get_adapter_name(transformer)
# In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
# otherwise loading LoRA weights will lead to an error
is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
if incompatible_keys is not None:
# check only for unexpected keys
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
logger.warning(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
# Offload back.
if is_model_cpu_offload:
_pipeline.enable_model_cpu_offload()
elif is_sequential_cpu_offload:
_pipeline.enable_sequential_cpu_offload()
# Unsafe code />
@classmethod
def save_lora_weights(
cls,
save_directory: Union[str, os.PathLike],
transformer_lora_layers: Dict[str, torch.nn.Module] = None,
is_main_process: bool = True,
weight_name: str = None,
save_function: Callable = None,
safe_serialization: bool = True,
):
r"""
Save the LoRA parameters corresponding to the UNet and text encoder.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to save LoRA parameters to. Will be created if it doesn't exist.
transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
State dict of the LoRA layers corresponding to the `transformer`.
is_main_process (`bool`, *optional*, defaults to `True`):
Whether the process calling this is the main process or not. Useful during distributed training and you
need to call this function on all processes. In this case, set `is_main_process=True` only on the main
process to avoid race conditions.
save_function (`Callable`):
The function to use to save the state dictionary. Useful during distributed training when you need to
replace `torch.save` with another method. Can be configured with the environment variable
`DIFFUSERS_SAVE_MODE`.
safe_serialization (`bool`, *optional*, defaults to `True`):
Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
"""
state_dict = {}
def pack_weights(layers, prefix):
layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
return layers_state_dict
if not transformer_lora_layers:
raise ValueError("You must pass `transformer_lora_layers`.")
if transformer_lora_layers:
state_dict.update(pack_weights(transformer_lora_layers, cls.transformer_name))
# Save the model
cls.write_lora_layers(
state_dict=state_dict,
save_directory=save_directory,
is_main_process=is_main_process,
weight_name=weight_name,
save_function=save_function,
safe_serialization=safe_serialization,
)
@staticmethod
def write_lora_layers(
state_dict: Dict[str, torch.Tensor],
save_directory: str,
is_main_process: bool,
weight_name: str,
save_function: Callable,
safe_serialization: bool,
):
if os.path.isfile(save_directory):
logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
return
if save_function is None:
if safe_serialization:
def save_function(weights, filename):
return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
else:
save_function = torch.save
os.makedirs(save_directory, exist_ok=True)
if weight_name is None:
if safe_serialization:
weight_name = LORA_WEIGHT_NAME_SAFE
else:
weight_name = LORA_WEIGHT_NAME
save_path = Path(save_directory, weight_name).as_posix()
save_function(state_dict, save_path)
logger.info(f"Model weights saved in {save_path}")
def unload_lora_weights(self):
"""
Unloads the LoRA parameters.
Examples:
```python
>>> # Assuming `pipeline` is already loaded with the LoRA parameters.
>>> pipeline.unload_lora_weights()
>>> ...
```
"""
transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
recurse_remove_peft_layers(transformer)
if hasattr(transformer, "peft_config"):
del transformer.peft_config
@classmethod
# Copied from diffusers.loaders.lora.LoraLoaderMixin._optionally_disable_offloading
def _optionally_disable_offloading(cls, _pipeline):
"""
Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
Args:
_pipeline (`DiffusionPipeline`):
The pipeline to disable offloading for.
Returns:
tuple:
A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
"""
is_model_cpu_offload = False
is_sequential_cpu_offload = False
if _pipeline is not None and _pipeline.hf_device_map is None:
for _, component in _pipeline.components.items():
if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
if not is_model_cpu_offload:
is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
if not is_sequential_cpu_offload:
is_sequential_cpu_offload = (
isinstance(component._hf_hook, AlignDevicesHook)
or hasattr(component._hf_hook, "hooks")
and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
)
logger.info(
"Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
)
remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
return (is_model_cpu_offload, is_sequential_cpu_offload)
def fuse_lora(
self,
fuse_transformer: bool = True,
lora_scale: float = 1.0,
safe_fusing: bool = False,
adapter_names: Optional[List[str]] = None,
):
r"""
Fuses the LoRA parameters into the original parameters of the corresponding blocks.
<Tip warning={true}>
This is an experimental API.
</Tip>
Args:
fuse_transformer (`bool`, defaults to `True`): Whether to fuse the transformer LoRA parameters.
lora_scale (`float`, defaults to 1.0):
Controls how much to influence the outputs with the LoRA parameters.
safe_fusing (`bool`, defaults to `False`):
Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
adapter_names (`List[str]`, *optional*):
Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
Example:
```py
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16
).to("cuda")
pipeline.load_lora_weights(
"nerijs/pixel-art-medium-128-v0.1",
weight_name="pixel-art-medium-128-v0.1.safetensors",
adapter_name="pixel",
)
pipeline.fuse_lora(lora_scale=0.7)
```
"""
if fuse_transformer:
self.num_fused_loras += 1
if fuse_transformer:
transformer = (
getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
)
transformer.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
def unfuse_lora(self, unfuse_transformer: bool = True):
r"""
Reverses the effect of
[`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
<Tip warning={true}>
This is an experimental API.
</Tip>
Args:
unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the transformer LoRA parameters.
"""
from peft.tuners.tuners_utils import BaseTunerLayer
transformer = getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer
if unfuse_transformer:
for module in transformer.modules():
if isinstance(module, BaseTunerLayer):
module.unmerge()
self.num_fused_loras -= 1