File size: 17,575 Bytes
302920f |
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 |
# Copyright 2025-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.
import warnings
from typing import Any, Optional, Union
import torch
import torch.nn as nn
from peft.tuners.tuners_utils import BaseTunerLayer, check_adapters_to_merge
from .config import RoadConfig, RoadVariant
class RoadLayer(BaseTunerLayer):
"""
Road layer.
Generally the idea of RoAD is to split the input vector into many 2D vectors and rotate each 2D vector with its own
2D rotation matrix. For additional flexibility, each rotation matrix is multiplied by a trainable scale.
when applied to vector R @ x each pair of elements of x is transformed like this: `y₀ = x₀ * α * cosθ - xₙ * α *
sinθ` and `yₙ = x₀ * α * sinθ + xₙ * α * cosθ`
The scales α and angles θ are learned for each pair of elements and, moreover, each of the 4 instances in the
rotation matrix may actually be different (when using variant 2 or 4).
Note that instead of using two consecutive elements x₀ x₁ we first split the whole vector into groups and pair
elements from the first with the second half of the same group, which allows for more efficient inference
implementation.
The adapter needs to only store the angles θ and scales α, rather than the full matrix R and the inference
implementation only needs to do elementwise vector multiplications.
For merging the weights, we make use of the following formula: R @ (W @ x + b) = (R @ W) @ x + R @ b. The lhs part
is how it is used in unmerged state (using efficient elementwise implementation instead of matrix multiplication)
and the rhs part is how it is used in merged state where (R @ W) becomes the new weight matrix and R @ b becomes
the new bias.
"""
adapter_layer_names: tuple[str, ...] = ("road_theta", "road_alpha")
other_param_names: tuple[str, ...] = ("variant", "group_size")
def __init__(self, base_layer: nn.Module, ephemeral_gpu_offload: bool = False, **kwargs) -> None:
self.base_layer = base_layer
self.variant = {}
self.group_size = {}
self.road_theta = nn.ParameterDict({})
self.road_alpha = nn.ParameterDict({})
self._disable_adapters = False
self.merged_adapters = []
base_layer = self.get_base_layer()
if isinstance(base_layer, nn.Linear):
in_features, out_features = base_layer.in_features, base_layer.out_features
else:
raise ValueError(f"Unsupported layer type '{type(base_layer)}' encountered, cannot apply RoAd adapter.")
self.in_features = in_features
self.out_features = out_features
@property
def _available_adapters(self) -> set[str]:
return {*self.road_theta}
def update_layer(
self,
adapter_name,
variant,
group_size,
init_weights,
inference_mode: bool = False,
):
self.variant[adapter_name] = variant
self.group_size[adapter_name] = group_size
if self.out_features % group_size != 0:
raise ValueError(
f"The out_features of the base layer must be divisible by group_size ({group_size}) when using RoadLayer."
)
# Actual trainable parameters
if variant == "road_1":
size = self.out_features // 2
elif variant == "road_2":
size = self.out_features
elif variant == "road_4":
size = self.out_features * 2
else:
raise ValueError(
f"Unsupported variant {variant} for RoadLayer. Supported variants are road_1, road_2, and road_4."
)
self.road_theta[adapter_name] = nn.Parameter(torch.empty(size))
self.road_alpha[adapter_name] = nn.Parameter(torch.empty(size))
self.reset_parameters(adapter_name, init_weights)
self._move_adapter_to_device_of_base_layer(adapter_name)
self.set_adapter(self.active_adapters, inference_mode=inference_mode)
def reset_parameters(self, adapter_name, init_weights):
if init_weights is False:
nn.init.normal_(self.road_theta[adapter_name].data, mean=0.0, std=0.5)
nn.init.normal_(self.road_alpha[adapter_name].data, mean=1.0, std=0.5)
return
nn.init.zeros_(self.road_theta[adapter_name].data)
nn.init.ones_(self.road_alpha[adapter_name].data)
class Linear(nn.Module, RoadLayer):
# Road implemented in a dense layer
def __init__(
self,
base_layer,
adapter_name: str,
variant: RoadVariant = "road_1",
group_size: int = 64,
init_weights: Union[bool, str] = True,
**kwargs,
) -> None:
super().__init__()
RoadLayer.__init__(self, base_layer, **kwargs)
self._active_adapter = adapter_name
self.update_layer(
adapter_name,
variant,
group_size,
init_weights=init_weights,
)
def _check_forward_args(self, x, *args, **kwargs):
"""Check if the arguments are compatible with the configs and state of the model"""
adapter_names = kwargs.get("adapter_names", None)
if adapter_names is None:
return
if len(x) != len(adapter_names):
msg = (
"Length of `adapter_names` should be the same as the number of inputs, but got "
f"{len(adapter_names)} and {len(x)} respectively."
)
raise ValueError(msg)
if self.merged:
# It is unclear what would be the right thing to do if users pass adapter_names and there are merged
# adapters. Therefore, it is better to raise an error in this case.
msg = "Cannot pass `adapter_names` when there are merged adapters, please call `unmerge_adapter` first."
raise ValueError(msg)
def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor:
self._check_forward_args(x, *args, **kwargs)
adapter_names = kwargs.pop("adapter_names", None)
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)
elif adapter_names is not None:
result = self._mixed_batch_forward(x, *args, adapter_names=adapter_names, **kwargs)
else:
result = self.base_layer(x, *args, **kwargs)
torch_result_dtype = result.dtype
for active_adapter in self.active_adapters:
if active_adapter not in self._available_adapters:
continue
result = self._cast_input_dtype(result, self.road_theta[active_adapter].dtype)
result = _apply_road(
self.variant[active_adapter],
self.group_size[active_adapter],
self.road_theta[active_adapter],
self.road_alpha[active_adapter],
result,
)
result = result.to(torch_result_dtype)
return result
def _mixed_batch_forward(
self, x: torch.Tensor, *args: Any, adapter_names: list[str], **kwargs: Any
) -> torch.Tensor:
# This is a special method that handles the case when users pass the argument `adapter_names`. This is an
# extra argument that allows mixing different adapters in the same batch at inference time.
result = self.base_layer(x, *args, **kwargs)
unique_adapters = set(adapter_names)
sub_batch_indices_list = []
for adapter in unique_adapters:
sub_batch_indices_list.append([index for index, item in enumerate(adapter_names) if item == adapter])
for i, active_adapter in enumerate(unique_adapters):
if active_adapter == "__base__":
continue
if active_adapter not in self._available_adapters:
continue
dtype = self.road_theta[active_adapter].data.dtype
# getting the sub-batch, passing it to Road layers and updating the corresponding indices of the linear
# layer output
sub_batch = result[sub_batch_indices_list[i]].to(dtype)
result[sub_batch_indices_list[i]] = _apply_road(
self.variant[active_adapter],
self.group_size[active_adapter],
self.road_theta[active_adapter],
self.road_alpha[active_adapter],
sub_batch,
)
return result
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._available_adapters:
base_layer = self.get_base_layer()
orig_dtype = base_layer.weight.dtype
road_R = _get_delta_weight(
self.variant[active_adapter],
self.group_size[active_adapter],
self.road_theta[active_adapter].data,
self.road_alpha[active_adapter].data,
)
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()
orig_weight = torch.matmul(road_R.to(orig_dtype), orig_weight)
if not torch.isfinite(orig_weight).all():
raise ValueError(
f"NaNs detected in the merged weights. The adapter {active_adapter} seems to be broken"
)
base_layer.weight.data = orig_weight.contiguous().to(orig_dtype)
if base_layer.bias is not None:
orig_bias = base_layer.bias.clone()
orig_bias = torch.matmul(road_R.to(orig_dtype), orig_bias)
if not torch.isfinite(orig_bias).all():
raise ValueError(
f"NaNs detected in the merged bias. The adapter {active_adapter} seems to be broken"
)
base_layer.bias.data = orig_bias.contiguous().to(orig_dtype)
else:
orig_weight = base_layer.weight.data
orig_weight = torch.matmul(road_R.to(orig_dtype), orig_weight)
base_layer.weight.data = orig_weight.contiguous().to(orig_dtype)
if base_layer.bias is not None:
orig_bias = base_layer.bias.data
orig_bias = torch.matmul(road_R.to(orig_dtype), orig_bias)
base_layer.bias.data = orig_bias.contiguous().to(orig_dtype)
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:
# Going in reverse order
active_adapter = self.merged_adapters.pop()
if active_adapter in self._available_adapters:
weight = self.get_base_layer().weight
orig_dtype = weight.dtype
road_R = _get_delta_weight(
self.variant[active_adapter],
self.group_size[active_adapter],
self.road_theta[active_adapter].data,
self.road_alpha[active_adapter].data,
)
# Since our matrix are not necessarily orthogonal we need inverse instead of transpose.
# In practice we expect this to basically always work since we start from block diagonal rotation matrix.
inv_road_R = torch.linalg.inv(road_R.to(torch.float32)).to(orig_dtype)
orig_weight = torch.matmul(inv_road_R, weight.data)
weight.data = orig_weight.contiguous()
if self.get_base_layer().bias is not None:
orig_bias = torch.matmul(inv_road_R, self.get_base_layer().bias.data)
self.get_base_layer().bias.data = orig_bias.contiguous()
def __repr__(self) -> str:
rep = super().__repr__()
return "road." + rep
def _get_delta_weight(variant: RoadVariant, group_size: int, road_theta: torch.Tensor, road_alpha: torch.Tensor):
first_col, second_col = _prepare_cols(variant, group_size, road_theta, road_alpha)
# To help understand the logic below consider how rope embeddings work
# here it is similar, but done in groups.
# https://discuss.huggingface.co/t/is-llama-rotary-embedding-implementation-correct/44509/3
# First column is simply put on the main diagonal
output_tensor = torch.diag(first_col)
# For second column we need to swap each half groups and add minus sign
size = second_col.shape[0]
swapped_second_col = second_col.reshape(-1, 2, group_size // 2)[:, [1, 0], :].flatten()
rotated_diag_second_col = torch.diag(swapped_second_col).reshape(-1, 2, group_size // 2, size)[:, [1, 0], :, :]
rotated_diag_second_col[:, 0, :, :] *= -1
rotated_diag_second_col = rotated_diag_second_col.reshape(size, size)
output_tensor += rotated_diag_second_col
return output_tensor
def _prepare_cols(
variant: RoadVariant, group_size: int, road_theta: torch.Tensor, road_alpha: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
# In inference mode, this can be cached
if variant == "road_1":
# In each group there are only group_size // 2 parameters that are reused
road_theta = road_theta.reshape(-1, group_size // 2).repeat_interleave(2, dim=0).flatten()
road_alpha = road_alpha.reshape(-1, group_size // 2).repeat_interleave(2, dim=0).flatten()
theta_cos = road_theta.cos()
theta_sin = road_theta.sin()
first_col = road_alpha * theta_cos
second_col = road_alpha * theta_sin
elif variant == "road_2":
# Each group has exactly group_size parameters
theta_cos = road_theta.cos()
theta_sin = road_theta.sin()
first_col = road_alpha * theta_cos
second_col = road_alpha * theta_sin
elif variant == "road_4":
# Each group has 2*group_size parameters, first half used for first column, second half for second column
road_theta = road_theta.reshape(-1, 2, group_size)
theta_cos = road_theta[:, 0, :].cos().flatten()
theta_sin = road_theta[:, 1, :].sin().flatten()
road_alpha = road_alpha.reshape(-1, 2, group_size)
alpha_1 = road_alpha[:, 0, :].flatten()
alpha_2 = road_alpha[:, 1, :].flatten()
first_col = alpha_1 * theta_cos
second_col = alpha_2 * theta_sin
else:
raise ValueError(
f"Unsupported variant {variant} for RoadLayer. Supported variants are road_1, road_2, and road_4."
)
return first_col, second_col
def _apply_road(
variant: RoadVariant, group_size: int, road_theta: torch.Tensor, road_alpha: torch.Tensor, x: torch.Tensor
):
first_col, second_col = _prepare_cols(variant, group_size, road_theta, road_alpha)
# Split in half groups and join back
# See equation 4 in the RoAD paper
x_grouped = x.reshape(-1, 2, group_size // 2)
x1 = x_grouped[:, 0, :]
x2 = x_grouped[:, 1, :]
rotate_half_x = torch.stack((-x2, x1), dim=1).reshape(x.shape)
result = x * first_col + rotate_half_x * second_col
return result
def dispatch_default(
target: torch.nn.Module,
adapter_name: str,
road_config: RoadConfig,
**kwargs,
) -> Optional[torch.nn.Module]:
new_module = None
if isinstance(target, BaseTunerLayer):
target_base_layer = target.get_base_layer()
else:
target_base_layer = target
if isinstance(target_base_layer, torch.nn.Linear):
new_module = Linear(target, adapter_name, **kwargs)
return new_module
|