Spaces:
Sleeping
Sleeping
File size: 17,395 Bytes
0aaa1f1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 |
# Copyright 2023 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.
# IMPORTANT: #
###################################################################
# ----------------------------------------------------------------#
# This file is deprecated and will be removed soon #
# (as soon as PEFT will become a required dependency for LoRA) #
# ----------------------------------------------------------------#
###################################################################
from typing import Optional, Tuple, Union
import torch
import torch.nn.functional as F
from torch import nn
from ..utils import logging
from ..utils.import_utils import is_transformers_available
if is_transformers_available():
from transformers import CLIPTextModel, CLIPTextModelWithProjection
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def text_encoder_attn_modules(text_encoder):
attn_modules = []
if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
for i, layer in enumerate(text_encoder.text_model.encoder.layers):
name = f"text_model.encoder.layers.{i}.self_attn"
mod = layer.self_attn
attn_modules.append((name, mod))
else:
raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
return attn_modules
def text_encoder_mlp_modules(text_encoder):
mlp_modules = []
if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
for i, layer in enumerate(text_encoder.text_model.encoder.layers):
mlp_mod = layer.mlp
name = f"text_model.encoder.layers.{i}.mlp"
mlp_modules.append((name, mlp_mod))
else:
raise ValueError(f"do not know how to get mlp modules for: {text_encoder.__class__.__name__}")
return mlp_modules
def adjust_lora_scale_text_encoder(text_encoder, lora_scale: float = 1.0):
for _, attn_module in text_encoder_attn_modules(text_encoder):
if isinstance(attn_module.q_proj, PatchedLoraProjection):
attn_module.q_proj.lora_scale = lora_scale
attn_module.k_proj.lora_scale = lora_scale
attn_module.v_proj.lora_scale = lora_scale
attn_module.out_proj.lora_scale = lora_scale
for _, mlp_module in text_encoder_mlp_modules(text_encoder):
if isinstance(mlp_module.fc1, PatchedLoraProjection):
mlp_module.fc1.lora_scale = lora_scale
mlp_module.fc2.lora_scale = lora_scale
class PatchedLoraProjection(torch.nn.Module):
def __init__(self, regular_linear_layer, lora_scale=1, network_alpha=None, rank=4, dtype=None):
super().__init__()
from ..models.lora import LoRALinearLayer
self.regular_linear_layer = regular_linear_layer
device = self.regular_linear_layer.weight.device
if dtype is None:
dtype = self.regular_linear_layer.weight.dtype
self.lora_linear_layer = LoRALinearLayer(
self.regular_linear_layer.in_features,
self.regular_linear_layer.out_features,
network_alpha=network_alpha,
device=device,
dtype=dtype,
rank=rank,
)
self.lora_scale = lora_scale
# overwrite PyTorch's `state_dict` to be sure that only the 'regular_linear_layer' weights are saved
# when saving the whole text encoder model and when LoRA is unloaded or fused
def state_dict(self, *args, destination=None, prefix="", keep_vars=False):
if self.lora_linear_layer is None:
return self.regular_linear_layer.state_dict(
*args, destination=destination, prefix=prefix, keep_vars=keep_vars
)
return super().state_dict(*args, destination=destination, prefix=prefix, keep_vars=keep_vars)
def _fuse_lora(self, lora_scale=1.0, safe_fusing=False):
if self.lora_linear_layer is None:
return
dtype, device = self.regular_linear_layer.weight.data.dtype, self.regular_linear_layer.weight.data.device
w_orig = self.regular_linear_layer.weight.data.float()
w_up = self.lora_linear_layer.up.weight.data.float()
w_down = self.lora_linear_layer.down.weight.data.float()
if self.lora_linear_layer.network_alpha is not None:
w_up = w_up * self.lora_linear_layer.network_alpha / self.lora_linear_layer.rank
fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
if safe_fusing and torch.isnan(fused_weight).any().item():
raise ValueError(
"This LoRA weight seems to be broken. "
f"Encountered NaN values when trying to fuse LoRA weights for {self}."
"LoRA weights will not be fused."
)
self.regular_linear_layer.weight.data = fused_weight.to(device=device, dtype=dtype)
# we can drop the lora layer now
self.lora_linear_layer = None
# offload the up and down matrices to CPU to not blow the memory
self.w_up = w_up.cpu()
self.w_down = w_down.cpu()
self.lora_scale = lora_scale
def _unfuse_lora(self):
if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
return
fused_weight = self.regular_linear_layer.weight.data
dtype, device = fused_weight.dtype, fused_weight.device
w_up = self.w_up.to(device=device).float()
w_down = self.w_down.to(device).float()
unfused_weight = fused_weight.float() - (self.lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
self.regular_linear_layer.weight.data = unfused_weight.to(device=device, dtype=dtype)
self.w_up = None
self.w_down = None
def forward(self, input):
if self.lora_scale is None:
self.lora_scale = 1.0
if self.lora_linear_layer is None:
return self.regular_linear_layer(input)
return self.regular_linear_layer(input) + (self.lora_scale * self.lora_linear_layer(input))
class LoRALinearLayer(nn.Module):
r"""
A linear layer that is used with LoRA.
Parameters:
in_features (`int`):
Number of input features.
out_features (`int`):
Number of output features.
rank (`int`, `optional`, defaults to 4):
The rank of the LoRA layer.
network_alpha (`float`, `optional`, defaults to `None`):
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. See
https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
device (`torch.device`, `optional`, defaults to `None`):
The device to use for the layer's weights.
dtype (`torch.dtype`, `optional`, defaults to `None`):
The dtype to use for the layer's weights.
"""
def __init__(
self,
in_features: int,
out_features: int,
rank: int = 4,
network_alpha: Optional[float] = None,
device: Optional[Union[torch.device, str]] = None,
dtype: Optional[torch.dtype] = None,
):
super().__init__()
self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
self.network_alpha = network_alpha
self.rank = rank
self.out_features = out_features
self.in_features = in_features
nn.init.normal_(self.down.weight, std=1 / rank)
nn.init.zeros_(self.up.weight)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
orig_dtype = hidden_states.dtype
dtype = self.down.weight.dtype
down_hidden_states = self.down(hidden_states.to(dtype))
up_hidden_states = self.up(down_hidden_states)
if self.network_alpha is not None:
up_hidden_states *= self.network_alpha / self.rank
return up_hidden_states.to(orig_dtype)
class LoRAConv2dLayer(nn.Module):
r"""
A convolutional layer that is used with LoRA.
Parameters:
in_features (`int`):
Number of input features.
out_features (`int`):
Number of output features.
rank (`int`, `optional`, defaults to 4):
The rank of the LoRA layer.
kernel_size (`int` or `tuple` of two `int`, `optional`, defaults to 1):
The kernel size of the convolution.
stride (`int` or `tuple` of two `int`, `optional`, defaults to 1):
The stride of the convolution.
padding (`int` or `tuple` of two `int` or `str`, `optional`, defaults to 0):
The padding of the convolution.
network_alpha (`float`, `optional`, defaults to `None`):
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. See
https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
"""
def __init__(
self,
in_features: int,
out_features: int,
rank: int = 4,
kernel_size: Union[int, Tuple[int, int]] = (1, 1),
stride: Union[int, Tuple[int, int]] = (1, 1),
padding: Union[int, Tuple[int, int], str] = 0,
network_alpha: Optional[float] = None,
):
super().__init__()
self.down = nn.Conv2d(in_features, rank, kernel_size=kernel_size, stride=stride, padding=padding, bias=False)
# according to the official kohya_ss trainer kernel_size are always fixed for the up layer
# # see: https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L129
self.up = nn.Conv2d(rank, out_features, kernel_size=(1, 1), stride=(1, 1), bias=False)
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
self.network_alpha = network_alpha
self.rank = rank
nn.init.normal_(self.down.weight, std=1 / rank)
nn.init.zeros_(self.up.weight)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
orig_dtype = hidden_states.dtype
dtype = self.down.weight.dtype
down_hidden_states = self.down(hidden_states.to(dtype))
up_hidden_states = self.up(down_hidden_states)
if self.network_alpha is not None:
up_hidden_states *= self.network_alpha / self.rank
return up_hidden_states.to(orig_dtype)
class LoRACompatibleConv(nn.Conv2d):
"""
A convolutional layer that can be used with LoRA.
"""
def __init__(self, *args, lora_layer: Optional[LoRAConv2dLayer] = None, **kwargs):
super().__init__(*args, **kwargs)
self.lora_layer = lora_layer
def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
self.lora_layer = lora_layer
def _fuse_lora(self, lora_scale: float = 1.0, safe_fusing: bool = False):
if self.lora_layer is None:
return
dtype, device = self.weight.data.dtype, self.weight.data.device
w_orig = self.weight.data.float()
w_up = self.lora_layer.up.weight.data.float()
w_down = self.lora_layer.down.weight.data.float()
if self.lora_layer.network_alpha is not None:
w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
fusion = torch.mm(w_up.flatten(start_dim=1), w_down.flatten(start_dim=1))
fusion = fusion.reshape((w_orig.shape))
fused_weight = w_orig + (lora_scale * fusion)
if safe_fusing and torch.isnan(fused_weight).any().item():
raise ValueError(
"This LoRA weight seems to be broken. "
f"Encountered NaN values when trying to fuse LoRA weights for {self}."
"LoRA weights will not be fused."
)
self.weight.data = fused_weight.to(device=device, dtype=dtype)
# we can drop the lora layer now
self.lora_layer = None
# offload the up and down matrices to CPU to not blow the memory
self.w_up = w_up.cpu()
self.w_down = w_down.cpu()
self._lora_scale = lora_scale
def _unfuse_lora(self):
if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
return
fused_weight = self.weight.data
dtype, device = fused_weight.data.dtype, fused_weight.data.device
self.w_up = self.w_up.to(device=device).float()
self.w_down = self.w_down.to(device).float()
fusion = torch.mm(self.w_up.flatten(start_dim=1), self.w_down.flatten(start_dim=1))
fusion = fusion.reshape((fused_weight.shape))
unfused_weight = fused_weight.float() - (self._lora_scale * fusion)
self.weight.data = unfused_weight.to(device=device, dtype=dtype)
self.w_up = None
self.w_down = None
def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
if self.lora_layer is None:
# make sure to the functional Conv2D function as otherwise torch.compile's graph will break
# see: https://github.com/huggingface/diffusers/pull/4315
return F.conv2d(
hidden_states, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups
)
else:
original_outputs = F.conv2d(
hidden_states, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups
)
return original_outputs + (scale * self.lora_layer(hidden_states))
class LoRACompatibleLinear(nn.Linear):
"""
A Linear layer that can be used with LoRA.
"""
def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
super().__init__(*args, **kwargs)
self.lora_layer = lora_layer
def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
self.lora_layer = lora_layer
def _fuse_lora(self, lora_scale: float = 1.0, safe_fusing: bool = False):
if self.lora_layer is None:
return
dtype, device = self.weight.data.dtype, self.weight.data.device
w_orig = self.weight.data.float()
w_up = self.lora_layer.up.weight.data.float()
w_down = self.lora_layer.down.weight.data.float()
if self.lora_layer.network_alpha is not None:
w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
if safe_fusing and torch.isnan(fused_weight).any().item():
raise ValueError(
"This LoRA weight seems to be broken. "
f"Encountered NaN values when trying to fuse LoRA weights for {self}."
"LoRA weights will not be fused."
)
self.weight.data = fused_weight.to(device=device, dtype=dtype)
# we can drop the lora layer now
self.lora_layer = None
# offload the up and down matrices to CPU to not blow the memory
self.w_up = w_up.cpu()
self.w_down = w_down.cpu()
self._lora_scale = lora_scale
def _unfuse_lora(self):
if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
return
fused_weight = self.weight.data
dtype, device = fused_weight.dtype, fused_weight.device
w_up = self.w_up.to(device=device).float()
w_down = self.w_down.to(device).float()
unfused_weight = fused_weight.float() - (self._lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
self.weight.data = unfused_weight.to(device=device, dtype=dtype)
self.w_up = None
self.w_down = None
def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
if self.lora_layer is None:
out = super().forward(hidden_states)
return out
else:
out = super().forward(hidden_states) + (scale * self.lora_layer(hidden_states))
return out
|