File size: 13,454 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 |
# 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 RandLora, since Randlora has some specific challenges due to the shared weights.
# These tests are copied from the test_vera.py file
import os
import pytest
import torch
from accelerate.utils.imports import is_bf16_available
from safetensors import safe_open
from torch import nn
from peft import PeftModel, RandLoraConfig, get_peft_model
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, 20, bias=bias) # lin1 and lin2 have same shape
self.lin2 = nn.Linear(20, 20, bias=bias)
self.lin3 = nn.Linear(20, 2, 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
# Tests copied from the TestVera class in test_vera.py.
# Changes to the code file should be reflected here.
class TestRandLora:
@pytest.fixture
def mlp(self):
torch.manual_seed(0)
model = MLP()
return model
@pytest.fixture
def mlp_same_prng(self, mlp):
torch.manual_seed(0)
config = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False)
# creates a default RandLora adapter
peft_model = get_peft_model(mlp, config)
config2 = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False)
peft_model.add_adapter("other", config2)
return peft_model
def test_multiple_adapters_same_prng_weights(self, mlp_same_prng):
# we can have multiple adapters with the same prng key, in which case the weights should be shared
assert (
mlp_same_prng.base_model.model.lin1.randlora_A["default"]
is mlp_same_prng.base_model.model.lin1.randlora_A["other"]
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_B["default"]
is mlp_same_prng.base_model.model.lin1.randlora_B["other"]
)
assert (
mlp_same_prng.base_model.model.lin2.randlora_A["default"]
is mlp_same_prng.base_model.model.lin2.randlora_A["other"]
)
assert (
mlp_same_prng.base_model.model.lin2.randlora_B["default"]
is mlp_same_prng.base_model.model.lin2.randlora_B["other"]
)
input = torch.randn(5, 10)
mlp_same_prng.set_adapter("default")
output_default = mlp_same_prng(input)
mlp_same_prng.set_adapter("other")
output_other = mlp_same_prng(input)
assert not torch.allclose(output_default, output_other, atol=1e-3, rtol=1e-3)
def test_multiple_adapters_different_prng_raises(self):
# we cannot have multiple adapters with different prng keys
model = MLP()
config = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False)
# creates a default RandLora adapter
peft_model = get_peft_model(model, config)
config2 = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False, projection_prng_key=123)
msg = (
r"RandLora PRNG initialisation key must be the same for all adapters. Got config.projection_prng_key=123 but "
r"previous config had 0"
)
with pytest.raises(ValueError, match=msg):
peft_model.add_adapter("other", config2)
def test_multiple_adapters_save_load_save_projection_true(self, mlp_same_prng, tmp_path):
# check saving and loading works with multiple adapters and saved projection weights
torch.manual_seed(0)
input = torch.randn(5, 10)
mlp_same_prng.set_adapter("default")
output_default = mlp_same_prng(input)
mlp_same_prng.set_adapter("other")
output_other = mlp_same_prng(input)
# sanity check
assert not torch.allclose(output_default, output_other, atol=1e-3, rtol=1e-3)
save_path = tmp_path / "randlora"
mlp_same_prng.save_pretrained(save_path)
assert os.path.exists(save_path / "adapter_config.json")
assert os.path.exists(save_path / "other" / "adapter_config.json")
torch.manual_seed(0)
mlp = MLP()
peft_model = PeftModel.from_pretrained(mlp, save_path)
peft_model.load_adapter(save_path / "other", "other")
peft_model.set_adapter("default")
output_default_loaded = peft_model(input)
peft_model.set_adapter("other")
output_other_loaded = peft_model(input)
assert torch.allclose(output_default, output_default_loaded, atol=1e-3, rtol=1e-3)
assert torch.allclose(output_other, output_other_loaded, atol=1e-3, rtol=1e-3)
def test_multiple_adapters_save_load_save_projection_false(self, mlp, tmp_path):
# check saving and loading works with multiple adapters without saved projection weights
torch.manual_seed(1)
config = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False, save_projection=False)
# creates a default RandLora adapter
peft_model = get_peft_model(mlp, config, adapter_name="first")
config2 = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False, save_projection=False)
peft_model.add_adapter("second", config2)
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 = tmp_path / "randlora"
peft_model.save_pretrained(save_path)
assert os.path.exists(save_path / "first" / "adapter_config.json")
assert os.path.exists(save_path / "second" / "adapter_config.json")
torch.manual_seed(0)
mlp = MLP()
peft_model = PeftModel.from_pretrained(mlp, save_path / "first", adapter_name="first")
peft_model.load_adapter(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, atol=1e-3, rtol=1e-3)
assert torch.allclose(output_second, output_second_loaded, atol=1e-3, rtol=1e-3)
def test_multiple_adapters_save_projection_true_contains_randlora_A_randlora_B(self, mlp_same_prng, tmp_path):
# check that the state_dicts don't contain the projection weights
save_path = tmp_path / "randlora"
mlp_same_prng.save_pretrained(save_path)
sd_default = {}
with safe_open(save_path / "adapter_model.safetensors", framework="pt", device="cpu") as f:
for key in f.keys():
sd_default[key] = f.get_tensor(key)
assert any("randlora_A" in key for key in sd_default)
assert any("randlora_B" in key for key in sd_default)
# default rank for RandLora is 32
assert sd_default["base_model.randlora_A"].shape == (32, 1, 20)
assert sd_default["base_model.randlora_B"].shape == (20, 1, 32)
sd_other = {}
with safe_open(save_path / "other" / "adapter_model.safetensors", framework="pt", device="cpu") as f:
for key in f.keys():
sd_other[key] = f.get_tensor(key)
assert any("randlora_A" in key for key in sd_other)
assert any("randlora_B" in key for key in sd_other)
assert sd_other["base_model.randlora_A"].shape == (32, 1, 20)
assert sd_other["base_model.randlora_B"].shape == (20, 1, 32)
def test_multiple_adapters_save_projection_false_contains_no_randlora_A_randlora_B(self, mlp, tmp_path):
torch.manual_seed(1)
config = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False, save_projection=False)
# creates a default RandLora adapter
peft_model = get_peft_model(mlp, config, adapter_name="first")
config2 = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False, save_projection=False)
peft_model.add_adapter("second", config2)
save_path = tmp_path / "randlora"
peft_model.save_pretrained(save_path)
sd_default = {}
with safe_open(save_path / "first" / "adapter_model.safetensors", framework="pt", device="cpu") as f:
for key in f.keys():
sd_default[key] = f.get_tensor(key)
assert not any("randlora_A" in key for key in sd_default)
assert not any("randlora_B" in key for key in sd_default)
sd_other = {}
with safe_open(save_path / "second" / "adapter_model.safetensors", framework="pt", device="cpu") as f:
for key in f.keys():
sd_other[key] = f.get_tensor(key)
assert not any("randlora_A" in key for key in sd_other)
assert not any("randlora_B" in key for key in sd_other)
def test_randlora_A_randlora_B_share_memory(self, mlp_same_prng):
randlora_A = mlp_same_prng.randlora_A["default"]
randlora_B = mlp_same_prng.randlora_B["default"]
# these tensors should share the same data
assert randlora_A.data_ptr() == mlp_same_prng.base_model.model.lin1.randlora_A["default"].data_ptr()
assert randlora_B.data_ptr() == mlp_same_prng.base_model.model.lin1.randlora_B["default"].data_ptr()
assert randlora_A.data_ptr() == mlp_same_prng.base_model.model.lin2.randlora_A["default"].data_ptr()
assert randlora_B.data_ptr() == mlp_same_prng.base_model.model.lin2.randlora_B["default"].data_ptr()
# sanity check: these tensors shouldn't share the same data
assert randlora_A.data_ptr() != randlora_B.data_ptr()
def test_randlora_lambda_dont_share_memory(self, mlp_same_prng):
# sanity check: these tensors shouldn't share the same data
assert (
mlp_same_prng.base_model.model.lin1.randlora_lambda["default"].data_ptr()
!= mlp_same_prng.base_model.model.lin1.randlora_lambda["other"].data_ptr()
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_lambda["default"].data_ptr()
!= mlp_same_prng.base_model.model.lin2.randlora_lambda["default"].data_ptr()
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_lambda["other"].data_ptr()
!= mlp_same_prng.base_model.model.lin2.randlora_lambda["other"].data_ptr()
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_gamma["default"].data_ptr()
!= mlp_same_prng.base_model.model.lin1.randlora_gamma["other"].data_ptr()
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_gamma["default"].data_ptr()
!= mlp_same_prng.base_model.model.lin2.randlora_gamma["default"].data_ptr()
)
assert (
mlp_same_prng.base_model.model.lin1.randlora_gamma["other"].data_ptr()
!= mlp_same_prng.base_model.model.lin2.randlora_gamma["other"].data_ptr()
)
def test_randlora_different_shapes(self, mlp):
config = RandLoraConfig(target_modules=["lin0", "lin3"], init_weights=False)
mlp_different_shapes = get_peft_model(mlp, config)
randlora_A = mlp_different_shapes.randlora_A["default"]
randlora_B = mlp_different_shapes.randlora_B["default"]
# sanity check
assert mlp.lin0.base_layer.weight.shape != mlp.lin3.base_layer.weight.shape
# lin0 has the largest output dimension, lin3 has the largest input dimension
# randlora_A should have the shape of (rank, largest_in), randlora_B should have the shape of (largest_out, rank)
assert randlora_A.shape == (config.r, 1, mlp.lin3.in_features)
assert randlora_B.shape == (mlp.lin0.out_features, 1, config.r)
# should not raise
input = torch.randn(5, 10)
mlp_different_shapes(input)
@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
def test_randlora_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 = RandLoraConfig(target_modules=["lin1", "lin2"], init_weights=False)
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
|