File size: 11,052 Bytes
43b7e92 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 |
# 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.
"""
PEFT utilities: Utilities related to peft library
"""
import collections
import importlib
from typing import Optional
from packaging import version
from .import_utils import is_peft_available, is_torch_available
if is_torch_available():
import torch
def recurse_remove_peft_layers(model):
r"""
Recursively replace all instances of `LoraLayer` with corresponding new layers in `model`.
"""
from peft.tuners.tuners_utils import BaseTunerLayer
has_base_layer_pattern = False
for module in model.modules():
if isinstance(module, BaseTunerLayer):
has_base_layer_pattern = hasattr(module, "base_layer")
break
if has_base_layer_pattern:
from peft.utils import _get_submodules
key_list = [key for key, _ in model.named_modules() if "lora" not in key]
for key in key_list:
try:
parent, target, target_name = _get_submodules(model, key)
except AttributeError:
continue
if hasattr(target, "base_layer"):
setattr(parent, target_name, target.get_base_layer())
else:
# This is for backwards compatibility with PEFT <= 0.6.2.
# TODO can be removed once that PEFT version is no longer supported.
from peft.tuners.lora import LoraLayer
for name, module in model.named_children():
if len(list(module.children())) > 0:
## compound module, go inside it
recurse_remove_peft_layers(module)
module_replaced = False
if isinstance(module, LoraLayer) and isinstance(module, torch.nn.Linear):
new_module = torch.nn.Linear(
module.in_features,
module.out_features,
bias=module.bias is not None,
).to(module.weight.device)
new_module.weight = module.weight
if module.bias is not None:
new_module.bias = module.bias
module_replaced = True
elif isinstance(module, LoraLayer) and isinstance(module, torch.nn.Conv2d):
new_module = torch.nn.Conv2d(
module.in_channels,
module.out_channels,
module.kernel_size,
module.stride,
module.padding,
module.dilation,
module.groups,
).to(module.weight.device)
new_module.weight = module.weight
if module.bias is not None:
new_module.bias = module.bias
module_replaced = True
if module_replaced:
setattr(model, name, new_module)
del module
if torch.cuda.is_available():
torch.cuda.empty_cache()
return model
def scale_lora_layers(model, weight):
"""
Adjust the weightage given to the LoRA layers of the model.
Args:
model (`torch.nn.Module`):
The model to scale.
weight (`float`):
The weight to be given to the LoRA layers.
"""
from peft.tuners.tuners_utils import BaseTunerLayer
if weight == 1.0:
return
for module in model.modules():
if isinstance(module, BaseTunerLayer):
module.scale_layer(weight)
def unscale_lora_layers(model, weight: Optional[float] = None):
"""
Removes the previously passed weight given to the LoRA layers of the model.
Args:
model (`torch.nn.Module`):
The model to scale.
weight (`float`, *optional*):
The weight to be given to the LoRA layers. If no scale is passed the scale of the lora layer will be
re-initialized to the correct value. If 0.0 is passed, we will re-initialize the scale with the correct
value.
"""
from peft.tuners.tuners_utils import BaseTunerLayer
if weight == 1.0:
return
for module in model.modules():
if isinstance(module, BaseTunerLayer):
if weight is not None and weight != 0:
module.unscale_layer(weight)
elif weight is not None and weight == 0:
for adapter_name in module.active_adapters:
# if weight == 0 unscale should re-set the scale to the original value.
module.set_scale(adapter_name, 1.0)
def get_peft_kwargs(rank_dict, network_alpha_dict, peft_state_dict, is_unet=True):
rank_pattern = {}
alpha_pattern = {}
r = lora_alpha = list(rank_dict.values())[0]
if len(set(rank_dict.values())) > 1:
# get the rank occuring the most number of times
r = collections.Counter(rank_dict.values()).most_common()[0][0]
# for modules with rank different from the most occuring rank, add it to the `rank_pattern`
rank_pattern = dict(filter(lambda x: x[1] != r, rank_dict.items()))
rank_pattern = {k.split(".lora_B.")[0]: v for k, v in rank_pattern.items()}
if network_alpha_dict is not None and len(network_alpha_dict) > 0:
if len(set(network_alpha_dict.values())) > 1:
# get the alpha occuring the most number of times
lora_alpha = collections.Counter(network_alpha_dict.values()).most_common()[0][0]
# for modules with alpha different from the most occuring alpha, add it to the `alpha_pattern`
alpha_pattern = dict(filter(lambda x: x[1] != lora_alpha, network_alpha_dict.items()))
if is_unet:
alpha_pattern = {
".".join(k.split(".lora_A.")[0].split(".")).replace(".alpha", ""): v
for k, v in alpha_pattern.items()
}
else:
alpha_pattern = {".".join(k.split(".down.")[0].split(".")[:-1]): v for k, v in alpha_pattern.items()}
else:
lora_alpha = set(network_alpha_dict.values()).pop()
# layer names without the Diffusers specific
target_modules = list({name.split(".lora")[0] for name in peft_state_dict.keys()})
use_dora = any("lora_magnitude_vector" in k for k in peft_state_dict)
lora_config_kwargs = {
"r": r,
"lora_alpha": lora_alpha,
"rank_pattern": rank_pattern,
"alpha_pattern": alpha_pattern,
"target_modules": target_modules,
"use_dora": use_dora,
}
return lora_config_kwargs
def get_adapter_name(model):
from peft.tuners.tuners_utils import BaseTunerLayer
for module in model.modules():
if isinstance(module, BaseTunerLayer):
return f"default_{len(module.r)}"
return "default_0"
def set_adapter_layers(model, enabled=True):
from peft.tuners.tuners_utils import BaseTunerLayer
for module in model.modules():
if isinstance(module, BaseTunerLayer):
# The recent version of PEFT needs to call `enable_adapters` instead
if hasattr(module, "enable_adapters"):
module.enable_adapters(enabled=enabled)
else:
module.disable_adapters = not enabled
def delete_adapter_layers(model, adapter_name):
from peft.tuners.tuners_utils import BaseTunerLayer
for module in model.modules():
if isinstance(module, BaseTunerLayer):
if hasattr(module, "delete_adapter"):
module.delete_adapter(adapter_name)
else:
raise ValueError(
"The version of PEFT you are using is not compatible, please use a version that is greater than 0.6.1"
)
# For transformers integration - we need to pop the adapter from the config
if getattr(model, "_hf_peft_config_loaded", False) and hasattr(model, "peft_config"):
model.peft_config.pop(adapter_name, None)
# In case all adapters are deleted, we need to delete the config
# and make sure to set the flag to False
if len(model.peft_config) == 0:
del model.peft_config
model._hf_peft_config_loaded = None
def set_weights_and_activate_adapters(model, adapter_names, weights):
from peft.tuners.tuners_utils import BaseTunerLayer
def get_module_weight(weight_for_adapter, module_name):
if not isinstance(weight_for_adapter, dict):
# If weight_for_adapter is a single number, always return it.
return weight_for_adapter
for layer_name, weight_ in weight_for_adapter.items():
if layer_name in module_name:
return weight_
parts = module_name.split(".")
# e.g. key = "down_blocks.1.attentions.0"
key = f"{parts[0]}.{parts[1]}.attentions.{parts[3]}"
block_weight = weight_for_adapter.get(key, 1.0)
return block_weight
# iterate over each adapter, make it active and set the corresponding scaling weight
for adapter_name, weight in zip(adapter_names, weights):
for module_name, module in model.named_modules():
if isinstance(module, BaseTunerLayer):
# For backward compatbility with previous PEFT versions
if hasattr(module, "set_adapter"):
module.set_adapter(adapter_name)
else:
module.active_adapter = adapter_name
module.set_scale(adapter_name, get_module_weight(weight, module_name))
# set multiple active adapters
for module in model.modules():
if isinstance(module, BaseTunerLayer):
# For backward compatbility with previous PEFT versions
if hasattr(module, "set_adapter"):
module.set_adapter(adapter_names)
else:
module.active_adapter = adapter_names
def check_peft_version(min_version: str) -> None:
r"""
Checks if the version of PEFT is compatible.
Args:
version (`str`):
The version of PEFT to check against.
"""
if not is_peft_available():
raise ValueError("PEFT is not installed. Please install it with `pip install peft`")
is_peft_version_compatible = version.parse(importlib.metadata.version("peft")) > version.parse(min_version)
if not is_peft_version_compatible:
raise ValueError(
f"The version of PEFT you are using is not compatible, please use a version that is greater"
f" than {min_version}"
)
|