Spaces:
Running
on
Zero
Running
on
Zero
File size: 2,719 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 |
# 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.
import torch
from .layer import AdaLoraLayer
class SVDQuantLinear(torch.nn.Module, AdaLoraLayer):
def __init__(
self,
base_layer,
adapter_name,
r: int = 0,
lora_alpha: int = 1,
lora_dropout: float = 0.0,
init_lora_weights: bool = True,
**kwargs,
) -> None:
super().__init__()
AdaLoraLayer.__init__(self, base_layer)
# self.base_layer and self.quant_linear_module are the same; we need the former for consistency and the latter
# for backwards compatibility
self.quant_linear_module = base_layer
self._active_adapter = adapter_name
self.update_layer(adapter_name, r, lora_alpha, lora_dropout, init_lora_weights)
def forward(self, x: torch.Tensor) -> torch.Tensor:
result = self.quant_linear_module(x)
if self.disable_adapters:
return result
for active_adapter in self.active_adapters:
if active_adapter not in self.lora_A.keys():
continue
lora_A = self.lora_A[active_adapter]
lora_B = self.lora_B[active_adapter]
lora_E = self.lora_E[active_adapter]
dropout = self.lora_dropout[active_adapter]
scaling = self.scaling[active_adapter]
ranknum = self.ranknum[active_adapter] + 1e-5
requires_conversion = not torch.is_autocast_enabled()
if requires_conversion:
expected_dtype = result.dtype
if x.dtype != torch.float32:
x = x.float()
output = (dropout(x) @ (lora_A * lora_E).T @ lora_B.T) * scaling / ranknum
# TODO: here, the dtype conversion is applied on the *whole expression*,
# not the intermediate result, unlike for SVDLinear8bitLT and
# SVDLinear4bit, is that correct?
if requires_conversion:
output = output.to(expected_dtype)
result += output
return result
def __repr__(self) -> str:
rep = super().__repr__()
return "adalora." + rep
|