Spaces:
Running
on
Zero
Running
on
Zero
File size: 40,922 Bytes
d711508 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 |
# 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.
from __future__ import annotations
import math
import os
import warnings
from contextlib import contextmanager
from typing import Any, Optional, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Function
from torch.utils.cpp_extension import load
from peft.tuners.tuners_utils import BaseTunerLayer, check_adapters_to_merge
_FBD_CUDA = None
# this function is a 1:1 copy from accelerate
@contextmanager
def patch_environment(**kwargs):
"""
A context manager that will add each keyword argument passed to `os.environ` and remove them when exiting.
Will convert the values in `kwargs` to strings and upper-case all the keys.
Example:
```python
>>> import os
>>> from accelerate.utils import patch_environment
>>> with patch_environment(FOO="bar"):
... print(os.environ["FOO"]) # prints "bar"
>>> print(os.environ["FOO"]) # raises KeyError
```
"""
existing_vars = {}
for key, value in kwargs.items():
key = key.upper()
if key in os.environ:
existing_vars[key] = os.environ[key]
os.environ[key] = str(value)
yield
for key in kwargs:
key = key.upper()
if key in existing_vars:
# restore previous value
os.environ[key] = existing_vars[key]
else:
os.environ.pop(key, None)
def get_fbd_cuda():
global _FBD_CUDA
if _FBD_CUDA is not None:
return _FBD_CUDA
curr_dir = os.path.dirname(__file__)
# need ninja to build the extension
try:
with patch_environment(CC="gcc", CXX="gcc"):
fbd_cuda = load(
name="fbd_cuda",
sources=[f"{curr_dir}/fbd/fbd_cuda.cpp", f"{curr_dir}/fbd/fbd_cuda_kernel.cu"],
verbose=True,
# build_directory='/tmp/' # for debugging
)
# extra_cuda_cflags = ['-std=c++14', '-ccbin=$$(which gcc-7)']) # cuda10.2 is not compatible with gcc9. Specify gcc 7
import fbd_cuda
except Exception as e:
warnings.warn(f"Failed to load the CUDA extension: {e}, check if ninja is available.")
warnings.warn("Setting boft_n_butterfly_factor to 1 to speed up the finetuning process.")
fbd_cuda = None
_FBD_CUDA = fbd_cuda
return _FBD_CUDA
class FastBlockDiag(Function):
"""
Implements a custom autograd Function for a fast block diagonal operation using CUDA.
This function is optimized for 4D tensors where the last two dimensions are equal, representing block diagonal
matrices for efficient computation on CUDA devices.
"""
@staticmethod
def forward(ctx, input):
"""
The forward method for FastBlockDiag.
Computes the block diagonal operation on the input tensor using a CUDA-optimized function. This method assumes
that the input is a 4D tensor where the last two dimensions are equal, which represent the blocks to be
diagonalized.
Parameters:
ctx: A context object that can be used to stash information for backward computation.
input (Tensor): The input tensor of shape (N, D, H, H), where `N` is the batch size,
`D` represents one additional dimension (In BOFT, the number of BOFT blocks), and `H` is the
size of the square blocks along the last two dimensions (In BOFT, the block size).
Returns:
Tensor: The resulting tensor after applying the block diagonal operation,
will have the shape (N, DxH, DxH).
"""
output = get_fbd_cuda().forward(input)[0]
ctx.save_for_backward(input)
return output
@staticmethod
def backward(ctx, grad_output):
(input,) = ctx.saved_tensors
grad_input = get_fbd_cuda().backward(grad_output, input)[0]
return grad_input
class MultiplicativeDropoutLayer(nn.Module):
"""
Implements the multiplicative dropout layer for BOFT.
"""
def __init__(self, p=0.0):
"""
Initializes the multiplicative dropout layer.
Parameters:
p (float): The probability of dropping out a block. Defaults to 0.0.
"""
super().__init__()
self.p = p
def forward(self, x):
"""
Applies multiplicative dropout to the input tensor.
Parameters:
x (Tensor): The input tensor of shape (N, D, H, H), where `N` is the batch size, `D` represents
one additional dimension (In BOFT, the number of BOFT blocks), and `H` is the size of the square
blocks along the last two dimensions (In BOFT, the block size).
"""
if self.training:
# Ensure the last two dimensions are the same
if x.shape[-1] != x.shape[-2]:
raise ValueError("The last two dimensions of input should be the same!")
N, D, H, _ = x.shape
# Randomly select one from N
n_random = torch.randint(0, N, (1,)).item()
# Create a mask with 1s for matrices to be replaced with identity and 0s otherwise
num_to_replace = int(self.p * D)
num_zeros = D - num_to_replace
# Generate a flat tensor with desired number of 1s and 0s
mask = torch.cat([torch.ones(num_to_replace, device=x.device), torch.zeros(num_zeros, device=x.device)])
# Shuffle and reshape the mask
mask = mask[torch.randperm(D)].view(1, D, 1, 1)
full_mask = torch.zeros(N, D, 1, 1, device=x.device)
full_mask[n_random] = mask
# Use the mask to combine original matrices and identity matrices
eye_matrix = torch.eye(H, device=x.device).repeat(N, D, 1, 1)
x = (1 - full_mask) * x + full_mask * eye_matrix
return x
class BOFTLayer(BaseTunerLayer):
"""
Implements the BOFT layer.
"""
# All names of layers that may contain (trainable) adapter weights
adapter_layer_names = ("boft_R", "boft_s")
# All names of other parameters that may contain adapter-related parameters
other_param_names = ("boft_block_size", "boft_block_num", "boft_dropout")
def __init__(self, base_layer: nn.Module, **kwargs) -> None:
"""
Initializes the BOFT layer.
Note, currently only support linear layer and convolutional layer, with further support for other layers to be
added soon.
Parameters:
base_layer: the pretrained model layer
"""
self.base_layer = base_layer
self.boft_block_size = {}
self.boft_block_num = {}
self.boft_dropout = nn.ModuleDict({})
self.boft_R = nn.ParameterDict({})
self.boft_s = nn.ParameterDict({})
# Mark the weight as unmerged
self._disable_adapters = False
self.merged_adapters = []
self.kwargs = kwargs
base_layer = self.get_base_layer()
if isinstance(base_layer, nn.Linear):
in_features, out_features = base_layer.in_features, base_layer.out_features
elif isinstance(base_layer, nn.Conv2d):
in_features, out_features = base_layer.in_channels, base_layer.out_channels
else:
raise ValueError(f"Unsupported layer type {type(base_layer)}")
self.in_features = in_features
self.out_features = out_features
def set_scale(self, adapter, scale):
if adapter not in self.scaling:
# Ignore the case where the adapter is not in the layer
return
warnings.warn("Scaling operation for BOFT not supported! Automatically set scale to 1.")
def scale_layer(self, scale: float) -> None:
if scale == 1:
return
for active_adapter in self.active_adapters:
if active_adapter not in self.boft_R.keys():
continue
warnings.warn("Scaling operation for BOFT not supported! Automatically set scale to 1.")
def unscale_layer(self, scale=None) -> None:
for active_adapter in self.active_adapters:
if active_adapter not in self.boft_R.keys():
continue
warnings.warn("Unscaling operation for BOFT not supported! Keeping scale to 1.")
def update_layer(
self, adapter_name, boft_block_size, boft_block_num, boft_n_butterfly_factor, boft_dropout, init_weights
):
"""
Update the linear layer with trainable BOFT weights. Override for other layer types.
"""
# to be consistent with the paper notation
boft_n_butterfly_factor = boft_n_butterfly_factor - 1
if boft_n_butterfly_factor < 0:
raise ValueError(
f"You can only specify boft_n_butterfly_factor {boft_n_butterfly_factor+1} to be a positive integer number."
)
# Initialize the MultiplicativeDropoutLayer for boft_dropout > 0.0.
if boft_dropout > 0.0:
boft_dropout_layer = MultiplicativeDropoutLayer(p=boft_dropout)
else:
boft_dropout_layer = nn.Identity()
self.boft_dropout.update(nn.ModuleDict({adapter_name: boft_dropout_layer}))
if boft_block_size == 0 and boft_block_num != 0:
if self.in_features % boft_block_num != 0:
raise ValueError(
f"in_features ({self.in_features}) must be divisible by boft_block_num ({boft_block_num})!"
)
if boft_n_butterfly_factor != 0:
if boft_n_butterfly_factor > int(math.log2(boft_block_num)):
raise ValueError(
f"Invalid combination of boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_num ({boft_block_num})!"
)
if boft_block_num % (2**boft_n_butterfly_factor) != 0:
raise ValueError(
f"boft_block_num ({boft_block_num}) must be a multiple of 2 raised to the power of boft_n_butterfly_factor ({boft_n_butterfly_factor+1})!"
)
boft_block_size = int(self.in_features // boft_block_num)
elif boft_block_size != 0 and boft_block_num == 0:
if self.in_features % boft_block_size != 0:
raise ValueError(
f"in_features ({self.in_features}) must be divisible by boft_block_size ({boft_block_size})!"
)
if boft_n_butterfly_factor != 0:
if self.in_features < (boft_block_size * (2**boft_n_butterfly_factor)):
raise ValueError(
f"Invalid combination of in_features ({self.in_features}), boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_size ({boft_block_size})!"
)
if self.in_features % (boft_block_size * (2**boft_n_butterfly_factor)) != 0:
raise ValueError(
f"Invalid combination of in_features ({self.in_features}), boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_size ({boft_block_size})!"
)
boft_block_num = int(self.in_features // boft_block_size)
else:
raise ValueError(
f"You can only specify either boft_block_size ({boft_block_size}) or boft_block_num ({boft_block_num}), but not both simultaneously or setting both"
"to be 0, because boft_block_size x boft_block_num != in_features."
)
# In OFT you can specify the number of blocks to be 1
if boft_n_butterfly_factor != 0:
if boft_block_num % 2 != 0:
raise ValueError(f"boft_block_num ({boft_block_num}) must be an even number!")
if boft_block_size % 2 != 0:
raise ValueError(f"boft_block_size ({boft_block_size}) must be an even number!")
# If there is no butterfly factor, then permutation matrix P will be an identity matrix.
P = torch.empty((boft_n_butterfly_factor + 1, self.in_features, self.in_features))
for i in range(boft_n_butterfly_factor + 1):
perm = self.block_butterfly_perm(
self.in_features, int(boft_block_num / (2 ** (i))), int(boft_block_size / 2), boft_n_butterfly_factor
)
perm_mat = self.perm2mat(perm)
P[i] = perm_mat
self.register_buffer("boft_P", P)
self.boft_R[adapter_name] = nn.Parameter(
torch.zeros(boft_n_butterfly_factor + 1, boft_block_num, boft_block_size, boft_block_size)
)
self.boft_s[adapter_name] = nn.Parameter(torch.ones(int(self.out_features), 1))
self.reset_boft_parameters(adapter_name, init_weights)
weight = getattr(self, "weight", None)
if weight is not None:
# the layer is already completely initialized, this is an update
if weight.dtype.is_floating_point or weight.dtype.is_complex:
self.to(weight.device, dtype=weight.dtype)
else:
self.to(weight.device)
# set the boft block size and number
self.boft_block_size[adapter_name] = boft_block_size
self.boft_block_num[adapter_name] = boft_block_num
self.set_adapter(self.active_adapters)
def reset_boft_parameters(self, adapter_name, init_weights):
"""
Reset the BOFT parameters.
"""
if init_weights is False:
nn.init.normal_(self.boft_R[adapter_name], mean=0.0, std=0.1)
nn.init.normal_(self.boft_s[adapter_name], mean=1.0, std=0.1)
return
if adapter_name in self.boft_R.keys():
if init_weights is True:
# initialize R to zero
nn.init.zeros_(self.boft_R[adapter_name])
nn.init.ones_(self.boft_s[adapter_name])
else:
raise ValueError(f"Unknown initialization {init_weights=}")
def perm2mat(self, indices):
"""
Convert permutation indices to permutation matrix.
Args:
indices: A list of indices representing the permutation.
"""
# Number of indices determines the size of the square matrix
n = len(indices)
# Initialize a matrix of zeros
perm_mat = torch.zeros((n, n))
# Set the 1s according to the indices
for i, idx in enumerate(indices):
perm_mat[i, idx] = 1
return perm_mat
def block_butterfly_perm(self, n, b, r=3, n_butterfly_factor=1):
"""
Define the permutation matrix for the block butterfly permutation.
Args:
n: size of the permutation matrix
b: desired number of blocks after multiplying with the permutation matrix
r: base block size of the block diagonal matrix, e.g. 2x2, 3x3, 5x5 etc.
"""
if n_butterfly_factor == 0:
return torch.arange(n)
if b * r * 2 > n:
raise ValueError("Invalid number of blocks!")
block_size = int(n // b)
indices = torch.arange(n)
def sort_block(b, r):
step = b / r
initial_order = torch.arange(b)
sorted_order = torch.empty(b, dtype=torch.long)
evens = torch.arange(0, step, 2)
odds = torch.arange(1, step, 2)
sorted_seq = torch.cat((evens, odds), dim=0)
for i, pos in enumerate(sorted_seq):
sorted_order[int(i * r) : int(i * r + r)] = initial_order[int(pos * r) : int(pos * r + r)]
return sorted_order
sorted_order = sort_block(block_size, r)
for i in range(0, n, block_size):
block_end = i + block_size
tmp_indices = indices[i:block_end]
indices[i:block_end] = tmp_indices[sorted_order]
return indices
def cayley_batch(self, data):
"""
Perform the Cayley parametrization on a batch of skew-symmetric matrices.
Args:
data: A batch of skew-symmetric matrices of shape (b, r, c).
"""
b, r, c = data.shape
# Ensure the input matrix is skew-symmetric
skew_mat = 0.5 * (data - data.transpose(1, 2))
id_mat = torch.eye(r, device=data.device).unsqueeze(0).expand(b, r, c)
# Perform the Cayley parametrization
Q = torch.linalg.solve(id_mat + skew_mat, id_mat - skew_mat, left=False)
return Q
class Linear(nn.Module, BOFTLayer):
"""
BOFT implemented in a dense layer.
"""
def __init__(
self,
base_layer,
adapter_name: str,
boft_block_size: int = 8,
boft_block_num: int = 0,
boft_n_butterfly_factor: int = 0,
boft_dropout: float = 0.1,
fan_in_fan_out: bool = False, # Set this to True if the layer to replace stores weight like (fan_in, fan_out)
init_weights: Union[bool, str] = True,
is_target_conv_1d_layer: bool = False,
**kwargs,
) -> None:
super().__init__()
BOFTLayer.__init__(self, base_layer, **kwargs)
self.fan_in_fan_out = fan_in_fan_out
self._active_adapter = adapter_name
# Attempt to load the CUDA extension during model initialization
if not get_fbd_cuda():
self.fbd_cuda_available = False
# If the CUDA extension is not available, set the butterfly factor to 1 to speed up the finetuning process
boft_n_butterfly_factor = 1
else:
self.fbd_cuda_available = True
self.update_layer(
adapter_name, boft_block_size, boft_block_num, boft_n_butterfly_factor, boft_dropout, init_weights
)
self.is_target_conv_1d_layer = is_target_conv_1d_layer
def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None:
"""
Merge the active adapter weights into the base weights
Args:
safe_merge (`bool`, *optional*):
If True, the merge operation will be performed in a copy of the original weights and check for NaNs
before merging the weights. This is useful if you want to check if the merge operation will produce
NaNs. Defaults to `False`.
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`.
"""
adapter_names = check_adapters_to_merge(self, adapter_names)
if not adapter_names:
# no adapter to merge
return
for active_adapter in adapter_names:
if active_adapter in self.boft_R.keys():
base_layer = self.get_base_layer()
if safe_merge:
# Note that safe_merge will be slower than the normal merge
# because of the copy operation.
orig_weight = base_layer.weight.data.clone()
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = torch.transpose(orig_weight, 0, 1)
orig_weight = torch.mm(butterfly_oft_mat, orig_weight)
orig_weight = torch.transpose(orig_weight, 0, 1)
orig_weight = orig_weight * boft_s
if not torch.isfinite(orig_weight).all():
raise ValueError(
f"NaNs detected in the merged weights. The adapter {active_adapter} seems to be broken"
)
self.base_layer.weight.data = orig_weight
else:
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = base_layer.weight.data.clone()
orig_weight = torch.transpose(orig_weight, 0, 1)
orig_weight = torch.mm(butterfly_oft_mat, orig_weight)
orig_weight = torch.transpose(orig_weight, 0, 1)
orig_weight = orig_weight * boft_s
self.base_layer.weight.data = orig_weight
self.merged_adapters.append(active_adapter)
def unmerge(self) -> None:
"""
This method unmerges all merged adapter layers from the base weights.
"""
if not self.merged:
warnings.warn("Already unmerged. Nothing to do.")
return
while len(self.merged_adapters) > 0:
active_adapter = self.merged_adapters.pop()
if active_adapter in self.boft_R.keys():
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = self.get_base_layer().weight.data.clone()
orig_weight = torch.transpose(orig_weight, 0, 1)
orig_weight = torch.mm(butterfly_oft_mat.t(), orig_weight)
orig_weight = torch.transpose(orig_weight, 0, 1)
self.get_base_layer().weight.data = orig_weight * (1 / boft_s)
def get_delta_weight(self, adapter) -> tuple[torch.Tensor, torch.Tensor]:
"""
Compute the delta weight for the given adapter.
Args:
adapter (str):
The name of the adapter for which the delta weight should be computed.
"""
boft_R = self.boft_R[adapter]
boft_s = self.boft_s[adapter]
N, D, H, _ = boft_R.shape
boft_R = boft_R.view(N * D, H, H)
orth_rotate_butterfly = self.cayley_batch(boft_R)
orth_rotate_butterfly = orth_rotate_butterfly.view(N, D, H, H)
if self.fbd_cuda_available:
block_diagonal_butterfly = FastBlockDiag.apply(orth_rotate_butterfly)
else:
orth_rotate_butterfly = orth_rotate_butterfly.squeeze(0)
block_diagonal_butterfly = torch.block_diag(*torch.unbind(orth_rotate_butterfly))
block_diagonal_butterfly = block_diagonal_butterfly.unsqueeze(0)
butterfly_oft_mat_batch = torch.bmm(block_diagonal_butterfly, self.boft_P.permute(0, 2, 1))
butterfly_oft_mat_batch = torch.bmm(self.boft_P, butterfly_oft_mat_batch)
butterfly_oft_mat = butterfly_oft_mat_batch[0]
for i in range(1, butterfly_oft_mat_batch.shape[0]):
butterfly_oft_mat = butterfly_oft_mat_batch[i] @ butterfly_oft_mat
return butterfly_oft_mat, boft_s
def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
previous_dtype = x.dtype
if self.disable_adapters:
if self.merged:
self.unmerge()
result = self.base_layer(x, *args, **kwargs)
elif self.merged:
result = self.base_layer(x, *args, **kwargs)
else:
boft_rotation = torch.eye(self.in_features, device=x.device)
boft_scale = torch.ones((int(self.out_features), 1), device=x.device)
for active_adapter in self.active_adapters:
if active_adapter not in self.boft_R.keys():
continue
boft_R = self.boft_R[active_adapter]
boft_s = self.boft_s[active_adapter]
dropout = self.boft_dropout[active_adapter]
N, D, H, _ = boft_R.shape
boft_R = boft_R.view(N * D, H, H)
orth_rotate_butterfly = self.cayley_batch(boft_R)
orth_rotate_butterfly = orth_rotate_butterfly.view(N, D, H, H)
orth_rotate_butterfly = dropout(orth_rotate_butterfly)
if self.fbd_cuda_available:
block_diagonal_butterfly = FastBlockDiag.apply(orth_rotate_butterfly)
else:
orth_rotate_butterfly = orth_rotate_butterfly.squeeze(0)
block_diagonal_butterfly = torch.block_diag(*torch.unbind(orth_rotate_butterfly))
block_diagonal_butterfly = block_diagonal_butterfly.unsqueeze(0)
butterfly_oft_mat_batch = torch.bmm(block_diagonal_butterfly, self.boft_P.permute(0, 2, 1))
butterfly_oft_mat_batch = torch.bmm(self.boft_P, butterfly_oft_mat_batch)
butterfly_oft_mat = butterfly_oft_mat_batch[0]
for i in range(1, butterfly_oft_mat_batch.shape[0]):
butterfly_oft_mat = butterfly_oft_mat_batch[i] @ butterfly_oft_mat
boft_rotation = butterfly_oft_mat @ boft_rotation
boft_scale = boft_s * boft_scale
x = x.to(self.get_base_layer().weight.data.dtype)
orig_weight = self.get_base_layer().weight.data
orig_weight = torch.transpose(orig_weight, 0, 1)
rotated_weight = torch.mm(boft_rotation, orig_weight)
rotated_weight = torch.transpose(rotated_weight, 0, 1)
scaled_rotated_weight = rotated_weight * boft_scale
result = F.linear(input=x, weight=scaled_rotated_weight, bias=self.base_layer.bias)
result = result.to(previous_dtype)
return result
def __repr__(self) -> str:
rep = super().__repr__()
return "boft." + rep
class Conv2d(nn.Module, BOFTLayer):
"""
BOFT implemented in a Conv2d layer.
"""
def __init__(
self,
base_layer: nn.Module,
adapter_name: str,
boft_block_size: int = 8,
boft_block_num: int = 0,
boft_n_butterfly_factor: int = 0,
boft_dropout: float = 0.1,
init_weights: Union[bool, str] = True,
**kwargs,
) -> None:
super().__init__()
BOFTLayer.__init__(self, base_layer)
self._active_adapter = adapter_name
# Attempt to load the CUDA extension during model initialization
if not get_fbd_cuda():
self.fbd_cuda_available = False
# If the CUDA extension is not available, set the butterfly factor to 1 to speed up the finetuning process
boft_n_butterfly_factor = 1
else:
self.fbd_cuda_available = True
self.update_layer(
adapter_name, boft_block_size, boft_block_num, boft_n_butterfly_factor, boft_dropout, init_weights
)
def update_layer(
self, adapter_name, boft_block_size, boft_block_num, boft_n_butterfly_factor, boft_dropout, init_weights
):
"""
Update the conv2d layer with trainable BOFT weights.
"""
# to be consistent with the paper notation
boft_n_butterfly_factor = boft_n_butterfly_factor - 1
if boft_n_butterfly_factor < 0:
raise ValueError(
f"You can only specify boft_n_butterfly_factor {boft_n_butterfly_factor+1} to be a positive integer number."
)
# Initialize the MultiplicativeDropoutLayer for boft_dropout > 0.0.
if boft_dropout > 0.0:
boft_dropout_layer = MultiplicativeDropoutLayer(p=boft_dropout)
else:
boft_dropout_layer = nn.Identity()
self.boft_dropout.update(nn.ModuleDict({adapter_name: boft_dropout_layer}))
# layer information from the base layer
base_layer = self.get_base_layer()
conv_filter_dim = self.in_features * base_layer.kernel_size[0] * base_layer.kernel_size[0]
# Initialize the BOFT parameters.
if not (boft_block_size != 0) ^ (boft_block_num != 0):
raise ValueError(
f"You can only specify either boft_block_size ({boft_block_size}) or boft_block_num ({boft_block_num}), but not both simultaneously, because boft_block_size x boft_block_num != in_features."
)
if boft_block_size == 0 and boft_block_num != 0:
if conv_filter_dim % boft_block_num != 0:
raise ValueError(
f"Convolutional kernel dimension ({conv_filter_dim}) must be divisible by boft_block_num ({boft_block_num})!"
)
if boft_n_butterfly_factor != 0:
if boft_n_butterfly_factor > int(math.log2(boft_block_num)):
raise ValueError(
f"Invalid combination of boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_num ({boft_block_num})!"
)
if boft_block_num % (2**boft_n_butterfly_factor) != 0:
raise ValueError(
f"boft_block_num ({boft_block_num}) must be a multiple of 2 raised to the power of boft_n_butterfly_factor ({boft_n_butterfly_factor+1})!"
)
boft_block_size = int(conv_filter_dim // boft_block_num)
elif boft_block_size != 0 and boft_block_num == 0:
if conv_filter_dim % boft_block_size != 0:
raise ValueError(
f"Convolutional kernel dimension ({conv_filter_dim}) must be divisible by boft_block_size ({boft_block_size})!"
)
if boft_n_butterfly_factor != 0:
if conv_filter_dim < (boft_block_size * (2**boft_n_butterfly_factor)):
raise ValueError(
f"Invalid combination of convolutional kernel dimension ({conv_filter_dim}), boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_size ({boft_block_size})!"
)
if conv_filter_dim % (boft_block_size * (2**boft_n_butterfly_factor)) != 0:
raise ValueError(
f"Invalid combination of convolutional kernel dimension ({conv_filter_dim}), boft_n_butterfly_factor ({boft_n_butterfly_factor+1}) and boft_block_size ({boft_block_size})!"
)
boft_block_num = int(conv_filter_dim // boft_block_size)
else:
raise ValueError("Unknown error!")
# In OFT you can specify the number of blocks to be 1
if boft_n_butterfly_factor != 0:
if boft_block_num % 2 != 0:
raise ValueError(f"boft_block_num ({boft_block_num}) must be an even number!")
if boft_block_size % 2 != 0:
raise ValueError(f"boft_block_size ({boft_block_size}) must be an even number!")
# If there is no butterfly factor, then permutation matrix P will be an identity matrix.
P = torch.empty((boft_n_butterfly_factor + 1, conv_filter_dim, conv_filter_dim))
for i in range(boft_n_butterfly_factor + 1):
perm = self.block_butterfly_perm(
conv_filter_dim, int(boft_block_num / (2 ** (i))), int(boft_block_size / 2), boft_n_butterfly_factor
)
perm_mat = self.perm2mat(perm)
P[i] = perm_mat
self.register_buffer("boft_P", P)
self.boft_R[adapter_name] = nn.Parameter(
torch.zeros(boft_n_butterfly_factor + 1, boft_block_num, boft_block_size, boft_block_size)
)
self.boft_s[adapter_name] = nn.Parameter(torch.ones(1, int(self.out_features)))
self.reset_boft_parameters(adapter_name, init_weights)
weight = getattr(self, "weight", None)
if weight is not None:
# the layer is already completely initialized, this is an update
if weight.dtype.is_floating_point or weight.dtype.is_complex:
self.to(weight.device, dtype=weight.dtype)
else:
self.to(weight.device)
self.set_adapter(self.active_adapters)
# set the boft block size and number
self.boft_block_size[adapter_name] = boft_block_size
self.boft_block_num[adapter_name] = boft_block_num
def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None:
"""
Merge the active adapter weights into the base weights
Args:
safe_merge (`bool`, *optional*):
If True, the merge operation will be performed in a copy of the original weights and check for NaNs
before merging the weights. This is useful if you want to check if the merge operation will produce
NaNs. Defaults to `False`.
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`.
"""
adapter_names = check_adapters_to_merge(self, adapter_names)
if not adapter_names:
# no adapter to merge
return
for active_adapter in adapter_names:
if active_adapter in self.boft_R.keys():
base_layer = self.get_base_layer()
if safe_merge:
# Note that safe_merge will be slower than the normal merge
# because of the copy operation.
orig_weight = base_layer.weight.data.clone()
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = orig_weight.view(
self.in_features * base_layer.kernel_size[0] * base_layer.kernel_size[0], self.out_features
)
orig_weight = torch.mm(butterfly_oft_mat, orig_weight)
orig_weight = orig_weight * boft_s
orig_weight = orig_weight.view(
self.out_features, self.in_features, base_layer.kernel_size[0], base_layer.kernel_size[0]
)
self.base_layer.weight.data = orig_weight
else:
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = base_layer.weight.data.clone()
orig_weight = orig_weight.view(
self.in_features * base_layer.kernel_size[0] * base_layer.kernel_size[0], self.out_features
)
orig_weight = torch.mm(butterfly_oft_mat, orig_weight)
orig_weight = orig_weight * boft_s
orig_weight = orig_weight.view(
self.out_features, self.in_features, base_layer.kernel_size[0], base_layer.kernel_size[0]
)
self.base_layer.weight.data = orig_weight
self.merged_adapters.append(active_adapter)
def unmerge(self) -> None:
"""
This method unmerges all merged adapter layers from the base weights.
"""
if not self.merged:
warnings.warn("Already unmerged. Nothing to do.")
return
while len(self.merged_adapters) > 0:
active_adapter = self.merged_adapters.pop()
if active_adapter in self.boft_R.keys():
butterfly_oft_mat, boft_s = self.get_delta_weight(active_adapter)
orig_weight = self.get_base_layer().weight.data.clone()
orig_weight = orig_weight.view(
self.in_features * self.get_base_layer().kernel_size[0] * self.get_base_layer().kernel_size[0],
self.out_features,
)
orig_weight = torch.mm(butterfly_oft_mat.t(), orig_weight)
orig_weight = orig_weight * (1 / boft_s)
orig_weight = orig_weight.view(
self.out_features,
self.in_features,
self.get_base_layer().kernel_size[0],
self.get_base_layer().kernel_size[0],
)
self.get_base_layer().weight.data = orig_weight
def get_delta_weight(self, adapter) -> tuple[torch.Tensor, torch.Tensor]:
"""
Compute the delta weight for the given adapter.
Args:
adapter (str):
The name of the adapter for which the delta weight should be computed.
"""
boft_R = self.boft_R[adapter]
boft_s = self.boft_s[adapter]
N, D, H, _ = boft_R.shape
boft_R = boft_R.view(N * D, H, H)
orth_rotate_butterfly = self.cayley_batch(boft_R)
orth_rotate_butterfly = orth_rotate_butterfly.view(N, D, H, H)
if self.fbd_cuda_available:
block_diagonal_butterfly = FastBlockDiag.apply(orth_rotate_butterfly)
else:
orth_rotate_butterfly = orth_rotate_butterfly.squeeze(0)
block_diagonal_butterfly = torch.block_diag(*torch.unbind(orth_rotate_butterfly))
block_diagonal_butterfly = block_diagonal_butterfly.unsqueeze(0)
butterfly_oft_mat_batch = torch.bmm(block_diagonal_butterfly, self.boft_P.permute(0, 2, 1))
butterfly_oft_mat_batch = torch.bmm(self.boft_P, butterfly_oft_mat_batch)
butterfly_oft_mat = butterfly_oft_mat_batch[0]
for i in range(1, butterfly_oft_mat_batch.shape[0]):
butterfly_oft_mat = butterfly_oft_mat_batch[i] @ butterfly_oft_mat
return butterfly_oft_mat, boft_s
def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
previous_dtype = x.dtype
if self.disable_adapters:
if self.merged:
self.unmerge()
result = self.base_layer(x, *args, **kwargs)
elif self.merged:
result = self.base_layer(x, *args, **kwargs)
else:
boft_rotation = torch.eye(
self.in_features * self.base_layer.kernel_size[0] * self.base_layer.kernel_size[0], device=x.device
)
boft_scale = torch.ones((1, int(self.out_features)), device=x.device)
for active_adapter in self.active_adapters:
if active_adapter not in self.boft_R.keys():
continue
boft_R = self.boft_R[active_adapter]
boft_s = self.boft_s[active_adapter]
dropout = self.boft_dropout[active_adapter]
N, D, H, _ = boft_R.shape
boft_R = boft_R.view(N * D, H, H)
orth_rotate_butterfly = self.cayley_batch(boft_R)
orth_rotate_butterfly = orth_rotate_butterfly.view(N, D, H, H)
orth_rotate_butterfly = dropout(orth_rotate_butterfly)
if self.fbd_cuda_available:
block_diagonal_butterfly = FastBlockDiag.apply(orth_rotate_butterfly)
else:
orth_rotate_butterfly = orth_rotate_butterfly.squeeze(0)
block_diagonal_butterfly = torch.block_diag(*torch.unbind(orth_rotate_butterfly))
block_diagonal_butterfly = block_diagonal_butterfly.unsqueeze(0)
butterfly_oft_mat_batch = torch.bmm(block_diagonal_butterfly, self.boft_P.permute(0, 2, 1))
butterfly_oft_mat_batch = torch.bmm(self.boft_P, butterfly_oft_mat_batch)
butterfly_oft_mat = butterfly_oft_mat_batch[0]
for i in range(1, butterfly_oft_mat_batch.shape[0]):
butterfly_oft_mat = butterfly_oft_mat_batch[i] @ butterfly_oft_mat
boft_rotation = butterfly_oft_mat @ boft_rotation
boft_scale = boft_s * boft_scale
x = x.to(self.base_layer.weight.data.dtype)
orig_weight = self.base_layer.weight.data
orig_weight = orig_weight.view(
self.in_features * self.base_layer.kernel_size[0] * self.base_layer.kernel_size[0],
self.out_features,
)
rotated_weight = torch.mm(boft_rotation, orig_weight)
scaled_rotated_weight = rotated_weight * boft_scale
scaled_rotated_weight = scaled_rotated_weight.view(
self.out_features, self.in_features, self.base_layer.kernel_size[0], self.base_layer.kernel_size[0]
)
result = F.conv2d(
input=x,
weight=scaled_rotated_weight,
bias=self.base_layer.bias,
padding=self.base_layer.padding[0],
stride=self.base_layer.stride[0],
)
result = result.to(previous_dtype)
return result
def __repr__(self) -> str:
rep = super().__repr__()
return "boft." + rep
|