File size: 13,652 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 |
# 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.
# This test file is for tests specific to SHiRA.
import os
import pytest
import torch
from accelerate.utils.imports import is_bf16_available
from torch import nn
from peft import PeftModel, ShiraConfig, get_peft_model
def custom_random_mask_function_with_custom_kwargs(custom_arg):
def mask_fn(base_layer, r):
"""
This mask function is similar to the random_mask provided in src/peft/tuners/shira/mask_functions.py except the
seed is derived from custom_kwargs. Please use this as an example to create your own custom sparse masks that
may use custom_kwargs. Remember, for a pretrained weight with shape m, n, mask_fn must return only one mask
(shape: m, n) which must be binary 0 or 1 with num_shira_parameters = r(m+n) for linear layers. Device and
dtype of mask must be same as base layer's weight's device and dtype.
"""
new_seed = custom_arg
shape = base_layer.weight.shape
num_shira_weights = r * (shape[0] + shape[1])
random_generator = torch.Generator()
random_generator.manual_seed(new_seed)
idx = (torch.randperm(base_layer.weight.numel(), generator=random_generator)[:num_shira_weights]).to(
base_layer.weight.device
)
val = torch.ones_like(idx.type(base_layer.weight.dtype))
mask = torch.zeros_like(base_layer.weight.view(1, -1))
mask = mask.scatter_(1, idx.unsqueeze(0), val.unsqueeze(0)).view(shape)
return mask
return mask_fn
class MLP(nn.Module):
def __init__(self, bias=True):
super().__init__()
self.relu = nn.ReLU()
self.lin0 = nn.Linear(10, 20, bias=bias)
self.lin1 = nn.Linear(20, 40, bias=bias) # lin1 and lin2 have same shape
self.lin2 = nn.Linear(40, 30, bias=bias)
self.lin3 = nn.Linear(30, 10, bias=bias)
self.sm = nn.LogSoftmax(dim=-1)
def forward(self, X):
X = self.lin0(X)
X = self.relu(X)
X = self.lin1(X)
X = self.relu(X)
X = self.lin2(X)
X = self.relu(X)
X = self.lin3(X)
X = self.sm(X)
return X
class TestShira:
@pytest.fixture
def mlp(self):
torch.manual_seed(0)
model = MLP()
return model
def test_mlp_single_adapter_shapes(self, mlp):
# torch.manual_seed(0)
r = 2
config = ShiraConfig(r=r, target_modules=["lin1", "lin2"])
# creates a default SHiRA adapter
peft_model = get_peft_model(mlp, config)
shira_weight1_size = peft_model.base_model.model.lin1.shira_weight["default"].shape[0]
shira_weight2_size = peft_model.base_model.model.lin2.shira_weight["default"].shape[0]
shira_indices1_size = peft_model.base_model.model.lin1.shira_indices["default"].shape[1]
shira_indices2_size = peft_model.base_model.model.lin2.shira_indices["default"].shape[1]
base_weight1_size = peft_model.base_model.model.lin1.base_layer.weight.shape
base_weight2_size = peft_model.base_model.model.lin2.base_layer.weight.shape
delta_weight1_shape = peft_model.base_model.model.lin1.get_delta_weight("default").shape
delta_weight2_shape = peft_model.base_model.model.lin2.get_delta_weight("default").shape
assert shira_weight1_size == r * (base_weight1_size[0] + base_weight1_size[1])
assert shira_weight2_size == r * (base_weight2_size[0] + base_weight2_size[1])
assert shira_weight1_size == shira_indices1_size
assert shira_weight2_size == shira_indices2_size
assert delta_weight1_shape == base_weight1_size
assert delta_weight2_shape == base_weight2_size
return peft_model
def test_multiple_adapters_save_load(self, mlp, tmp_path):
# check saving and loading works with multiple adapters
# note, the random seeds in the below two configs are not the default values.
# so it will lead to different random sparse masks between saving and loading.
# our goal is to make sure that loaded indices are exactly the same as the saved indices regardless of what initial random mask gets generated.
# we will also make sure that parameters are saved and loaded correctly, and the output remains the same.
config = ShiraConfig(r=2, target_modules=["lin1", "lin2"], random_seed=56)
# creates a default SHiRA adapter
peft_model = get_peft_model(mlp, config, adapter_name="first")
config2 = ShiraConfig(r=3, target_modules=["lin1", "lin2", "lin3"], random_seed=67)
peft_model.add_adapter("second", config2)
assert torch.all(peft_model.base_model.model.lin1.shira_weight["first"] == 0)
assert torch.all(peft_model.base_model.model.lin2.shira_weight["first"] == 0)
assert torch.all(peft_model.base_model.model.lin1.shira_weight["second"] == 0)
assert torch.all(peft_model.base_model.model.lin2.shira_weight["second"] == 0)
assert torch.all(peft_model.base_model.model.lin3.shira_weight["second"] == 0)
shira_assign_val1_f = torch.randn_like(peft_model.base_model.model.lin1.shira_weight["first"])
peft_model.base_model.model.lin1.shira_weight["first"] = shira_assign_val1_f
shira_indices1_f = peft_model.base_model.model.lin1.shira_indices["first"]
shira_assign_val2_f = torch.randn_like(peft_model.base_model.model.lin2.shira_weight["first"])
peft_model.base_model.model.lin2.shira_weight["first"] = shira_assign_val2_f
shira_indices2_f = peft_model.base_model.model.lin2.shira_indices["first"]
shira_assign_val1_s = torch.randn_like(peft_model.base_model.model.lin1.shira_weight["second"])
peft_model.base_model.model.lin1.shira_weight["second"] = shira_assign_val1_s
shira_indices1_s = peft_model.base_model.model.lin1.shira_indices["second"]
shira_assign_val2_s = torch.randn_like(peft_model.base_model.model.lin2.shira_weight["second"])
peft_model.base_model.model.lin2.shira_weight["second"] = shira_assign_val2_s
shira_indices2_s = peft_model.base_model.model.lin2.shira_indices["second"]
shira_assign_val3_s = torch.randn_like(peft_model.base_model.model.lin3.shira_weight["second"])
peft_model.base_model.model.lin3.shira_weight["second"] = shira_assign_val3_s
shira_indices3_s = peft_model.base_model.model.lin3.shira_indices["second"]
input = torch.randn(5, 10)
peft_model.set_adapter("first")
output_first = peft_model(input)
peft_model.set_adapter("second")
output_second = peft_model(input)
# sanity check
assert not torch.allclose(output_first, output_second, atol=1e-3, rtol=1e-3)
save_path = os.path.join(tmp_path, "shira")
peft_model.save_pretrained(save_path)
assert os.path.exists(os.path.join(save_path, "first", "adapter_config.json"))
assert os.path.exists(os.path.join(save_path, "second", "adapter_config.json"))
del peft_model
torch.manual_seed(0)
mlp = MLP()
peft_model = PeftModel.from_pretrained(mlp, os.path.join(save_path, "first"), adapter_name="first")
peft_model.load_adapter(os.path.join(save_path, "second"), "second")
peft_model.set_adapter("first")
output_first_loaded = peft_model(input)
peft_model.set_adapter("second")
output_second_loaded = peft_model(input)
assert torch.allclose(output_first, output_first_loaded)
assert torch.allclose(output_second, output_second_loaded)
assert torch.all(shira_assign_val1_f == peft_model.base_model.model.lin1.shira_weight["first"])
assert torch.all(shira_assign_val2_f == peft_model.base_model.model.lin2.shira_weight["first"])
assert torch.all(shira_indices1_f == peft_model.base_model.model.lin1.shira_indices["first"])
assert torch.all(shira_indices2_f == peft_model.base_model.model.lin2.shira_indices["first"])
assert torch.all(shira_assign_val1_s == peft_model.base_model.model.lin1.shira_weight["second"])
assert torch.all(shira_assign_val2_s == peft_model.base_model.model.lin2.shira_weight["second"])
assert torch.all(shira_assign_val3_s == peft_model.base_model.model.lin3.shira_weight["second"])
assert torch.all(shira_indices1_s == peft_model.base_model.model.lin1.shira_indices["second"])
assert torch.all(shira_indices2_s == peft_model.base_model.model.lin2.shira_indices["second"])
assert torch.all(shira_indices3_s == peft_model.base_model.model.lin3.shira_indices["second"])
return peft_model
def test_save_load_custom_mask_function(self, mlp, tmp_path):
# we want to see if saving and loading works when a custom mask is involved
config = ShiraConfig(r=2, mask_type="custom", target_modules=["lin1", "lin2"], init_weights=False)
custom_arg = 120
custom_mask_fn = custom_random_mask_function_with_custom_kwargs(custom_arg)
config.mask_fn = custom_mask_fn
# create a custom mask SHiRA adapter
peft_model = get_peft_model(mlp, config, adapter_name="first")
shira_assign_val1_f = peft_model.base_model.model.lin1.shira_weight["first"]
shira_indices1_f = peft_model.base_model.model.lin1.shira_indices["first"]
shira_assign_val2_f = peft_model.base_model.model.lin2.shira_weight["first"]
shira_indices2_f = peft_model.base_model.model.lin2.shira_indices["first"]
input = torch.randn(5, 10)
peft_model.set_adapter("first")
output_first = peft_model(input)
save_path = os.path.join(tmp_path, "shira")
peft_model.save_pretrained(save_path)
assert os.path.exists(os.path.join(save_path, "first", "adapter_config.json"))
del peft_model
torch.manual_seed(0)
mlp = MLP()
peft_model = PeftModel.from_pretrained(mlp, os.path.join(save_path, "first"), adapter_name="first")
peft_model.set_adapter("first")
output_first_loaded = peft_model(input)
assert torch.allclose(output_first, output_first_loaded)
assert torch.all(shira_assign_val1_f == peft_model.base_model.model.lin1.shira_weight["first"])
assert torch.all(shira_assign_val2_f == peft_model.base_model.model.lin2.shira_weight["first"])
assert torch.all(shira_indices1_f == peft_model.base_model.model.lin1.shira_indices["first"])
assert torch.all(shira_indices2_f == peft_model.base_model.model.lin2.shira_indices["first"])
return peft_model
def test_save_load_default_random_mask_with_seed_function(self, mlp, tmp_path):
# we want to see if saving and loading works when a random mask is involved but the random seed is fixed.
config = ShiraConfig(r=2, target_modules=["lin1", "lin2"], random_seed=567, init_weights=False)
# create a custom mask SHiRA adapter
peft_model = get_peft_model(mlp, config, adapter_name="first")
shira_assign_val1_f = peft_model.base_model.model.lin1.shira_weight["first"]
shira_indices1_f = peft_model.base_model.model.lin1.shira_indices["first"]
shira_assign_val2_f = peft_model.base_model.model.lin2.shira_weight["first"]
shira_indices2_f = peft_model.base_model.model.lin2.shira_indices["first"]
input = torch.randn(5, 10)
peft_model.set_adapter("first")
output_first = peft_model(input)
save_path = os.path.join(tmp_path, "shira")
peft_model.save_pretrained(save_path)
assert os.path.exists(os.path.join(save_path, "first", "adapter_config.json"))
del peft_model
torch.manual_seed(0)
mlp = MLP()
peft_model = PeftModel.from_pretrained(mlp, os.path.join(save_path, "first"), adapter_name="first")
peft_model.set_adapter("first")
output_first_loaded = peft_model(input)
assert torch.allclose(output_first, output_first_loaded)
assert torch.all(shira_assign_val1_f == peft_model.base_model.model.lin1.shira_weight["first"])
assert torch.all(shira_assign_val2_f == peft_model.base_model.model.lin2.shira_weight["first"])
assert torch.all(shira_indices1_f == peft_model.base_model.model.lin1.shira_indices["first"])
assert torch.all(shira_indices2_f == peft_model.base_model.model.lin2.shira_indices["first"])
return peft_model
@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
def test_shira_dtypes(self, dtype):
if dtype == torch.bfloat16:
# skip if bf16 is not supported on hardware, see #1872
if not is_bf16_available():
pytest.skip("bfloat16 not supported on this system, skipping the test")
model = MLP().to(dtype)
config = ShiraConfig(r=2, target_modules=["lin1", "lin2"])
peft_model = get_peft_model(model, config)
inputs = torch.randn(5, 10).to(dtype)
output = peft_model(inputs) # should not raise
assert output.dtype == dtype
|