File size: 19,463 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 |
# 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 os
import tempfile
import pytest
import torch
from torch.testing import assert_close
from transformers import AutoModelForCausalLM
from peft import get_peft_model
from peft.peft_model import PeftModel
from peft.tuners.adaption_prompt import AdaptionPromptConfig
from peft.utils import infer_device
from peft.utils.other import prepare_model_for_kbit_training
from peft.utils.save_and_load import get_peft_model_state_dict
MODELS_TO_TEST = [
"hf-internal-testing/tiny-random-gpt2",
"trl-internal-testing/tiny-random-LlamaForCausalLM",
"hf-internal-testing/tiny-random-MistralForCausalLM",
]
class TestAdaptionPrompt:
"""
Tests for the AdaptionPrompt model.
Some of these tests were adapted from `test_peft_model.py` (which has been refactored since), but since we haven't
checked in the test checkpoints for Llama into `hf-internal-testing`, we separate them for now.
"""
transformers_class = AutoModelForCausalLM
torch_device = infer_device()
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_attributes(self, model_id):
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=1, adapter_len=4)
model = get_peft_model(model, config)
assert hasattr(model, "save_pretrained")
assert hasattr(model, "from_pretrained")
assert hasattr(model, "push_to_hub")
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_prepare_for_training(self, model_id):
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=1, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
model = model.to(self.torch_device)
dummy_input = torch.LongTensor([[1, 1, 1]]).to(self.torch_device)
dummy_output = model.get_input_embeddings()(dummy_input)
assert not dummy_output.requires_grad
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_prepare_for_int8_training(self, model_id):
model = self.transformers_class.from_pretrained(model_id)
model = prepare_model_for_kbit_training(model)
model = model.to(self.torch_device)
for param in model.parameters():
assert not param.requires_grad
config = AdaptionPromptConfig(adapter_layers=1, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
# For backward compatibility
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
dummy_input = torch.LongTensor([[1, 1, 1]]).to(self.torch_device)
dummy_output = model.get_input_embeddings()(dummy_input)
assert dummy_output.requires_grad
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_save_pretrained_regression(self, model_id):
seed = 420
torch.manual_seed(seed)
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname, safe_serialization=False)
torch.manual_seed(seed)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
# check if the state dicts are equal
state_dict = get_peft_model_state_dict(model)
state_dict_from_pretrained = get_peft_model_state_dict(model_from_pretrained)
# check if same keys
assert state_dict.keys() == state_dict_from_pretrained.keys()
# Check that the number of saved parameters is 4 -- 2 layers of (tokens and gate).
assert len(state_dict) == 4
# check if tensors equal
for key in state_dict.keys():
assert torch.allclose(
state_dict[key].to(self.torch_device), state_dict_from_pretrained[key].to(self.torch_device)
)
# check if `adapter_model.bin` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_model.bin"))
# check if `adapter_config.json` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_config.json"))
# check if `model.safetensors` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "model.safetensors"))
# check if `config.json` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "config.json"))
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_save_pretrained(self, model_id):
seed = 420
torch.manual_seed(seed)
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
torch.manual_seed(seed)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
# check if the state dicts are equal
state_dict = get_peft_model_state_dict(model)
state_dict_from_pretrained = get_peft_model_state_dict(model_from_pretrained)
# check if same keys
assert state_dict.keys() == state_dict_from_pretrained.keys()
# Check that the number of saved parameters is 4 -- 2 layers of (tokens and gate).
assert len(state_dict) == 4
# check if tensors equal
for key in state_dict.keys():
assert torch.allclose(
state_dict[key].to(self.torch_device), state_dict_from_pretrained[key].to(self.torch_device)
)
# check if `adapter_model.bin` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_model.safetensors"))
# check if `adapter_config.json` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_config.json"))
# check if `model.safetensors` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "model.safetensors"))
# check if `config.json` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "config.json"))
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_save_pretrained_selected_adapters(self, model_id):
seed = 420
torch.manual_seed(seed)
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
model = model.to(self.torch_device)
new_adapter_config = AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
model.add_adapter("new_adapter", new_adapter_config)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
torch.manual_seed(seed)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
model_from_pretrained.load_adapter(tmp_dirname, "new_adapter")
# check if the state dicts are equal
state_dict = get_peft_model_state_dict(model)
state_dict_from_pretrained = get_peft_model_state_dict(model_from_pretrained)
# check if same keys
assert state_dict.keys() == state_dict_from_pretrained.keys()
# Check that the number of saved parameters is 4 -- 2 layers of (tokens and gate).
assert len(state_dict) == 4
# check if tensors equal
for key in state_dict.keys():
assert torch.allclose(
state_dict[key].to(self.torch_device), state_dict_from_pretrained[key].to(self.torch_device)
)
# check if `adapter_model.bin` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_model.safetensors"))
# check if `adapter_config.json` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_config.json"))
# check if `model.safetensors` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "model.safetensors"))
# check if `config.json` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "config.json"))
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_generate(self, model_id):
model = self.transformers_class.from_pretrained(model_id)
config = AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config)
model = model.to(self.torch_device)
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
attention_mask = torch.LongTensor([[1, 1, 1], [1, 0, 1]]).to(self.torch_device)
# check if `generate` works
_ = model.generate(input_ids=input_ids, attention_mask=attention_mask)
# check if `generate` works if positional arguments are passed
_ = model.generate(input_ids, attention_mask=attention_mask)
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_sequence_adapter_ops(self, model_id):
"""Test sequence of adapter operations."""
# Test input data.
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
target_ids = torch.LongTensor([[0, 0, 0], [0, 0, 0]]).to(self.torch_device)
attention_mask = torch.LongTensor([[1, 1, 1], [1, 0, 1]]).to(self.torch_device)
# Create original llama model.
original = self.transformers_class.from_pretrained(model_id)
original = original.to(self.torch_device)
original_before = original(input_ids=input_ids, attention_mask=attention_mask)
# Get AdaptionPrompt model.
adapted = get_peft_model(
original, AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
)
adapted = adapted.to(self.torch_device)
default_before = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
# Test zero-init: The logits should be exactly the same.
assert_close(original_before.logits, default_before.logits, rtol=0, atol=0)
# Single fine-tuning step on "default" adapter.
optimizer = torch.optim.SGD(adapted.parameters(), lr=1)
optimizer.zero_grad()
default_before.loss.backward()
optimizer.step()
# Test that the output changed.
default_after = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert not torch.allclose(default_before.logits, default_after.logits)
with adapted.disable_adapter():
# Test that the output is the same as the original output.
default_disabled = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(original_before.logits, default_disabled.logits, rtol=0, atol=0)
# Add new adapter 1.
adapted.add_adapter("adapter 1", AdaptionPromptConfig(adapter_layers=2, adapter_len=8, task_type="CAUSAL_LM"))
# Test zero-init
adapter_1_before = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(original_before.logits, adapter_1_before.logits, rtol=0, atol=0)
# Single fine-tuning step on adapter 1.
optimizer = torch.optim.SGD(adapted.parameters(), lr=1)
optimizer.zero_grad()
adapter_1_before.loss.backward()
optimizer.step()
# Test that adapter 1 output changed.
adapter_1_after = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert not torch.allclose(adapter_1_before.logits, adapter_1_after.logits)
assert not torch.allclose(original_before.logits, adapter_1_after.logits)
assert not torch.allclose(default_after.logits, adapter_1_after.logits)
with adapted.disable_adapter():
# Test that the output is the same as the original output.
adapter_1_disabled = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(original_before.logits, adapter_1_disabled.logits, rtol=0, atol=0)
# Set adapter back to default.
adapted.set_adapter("default")
# Test that the output is the same as the default output after training.
default_after_set = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(default_after.logits, default_after_set.logits, rtol=0, atol=0)
assert not torch.allclose(original_before.logits, default_after_set.logits)
assert not torch.allclose(adapter_1_after.logits, default_after_set.logits)
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_add_and_set_while_disabled(self, model_id):
"""Test that adding and setting adapters while disabled works as intended."""
# Test input data.
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
target_ids = torch.LongTensor([[0, 0, 0], [0, 0, 0]]).to(self.torch_device)
attention_mask = torch.LongTensor([[1, 1, 1], [1, 0, 1]]).to(self.torch_device)
# Create original llama model.
original = self.transformers_class.from_pretrained(model_id)
original = original.to(self.torch_device)
original_before = original(input_ids=input_ids, attention_mask=attention_mask)
# Get AdaptionPrompt model.
adapted = get_peft_model(
original, AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
)
adapted = adapted.to(self.torch_device)
with adapted.disable_adapter():
adapted.add_adapter(
"adapter 1", AdaptionPromptConfig(adapter_layers=2, adapter_len=8, task_type="CAUSAL_LM")
)
# Test that the output is the same as the original output.
adapter_1_before = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(original_before.logits, adapter_1_before.logits, rtol=0, atol=0)
# Single fine-tuning step on adapter 1.
optimizer = torch.optim.SGD(adapted.parameters(), lr=1)
optimizer.zero_grad()
adapter_1_before.loss.backward()
optimizer.step()
# Test that adapter 1 output changed.
adapter_1_after = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert not torch.allclose(original_before.logits, adapter_1_after.logits)
adapted.set_adapter("default")
with adapted.disable_adapter():
adapted.set_adapter("adapter 1")
# Test that adapter 1 is active again.
adapter_1_after_set = adapted(input_ids=input_ids, attention_mask=attention_mask, labels=target_ids)
assert_close(adapter_1_after.logits, adapter_1_after_set.logits, rtol=0, atol=0)
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_use_cache(self, model_id):
"""Test that AdaptionPrompt works when Llama config use_cache=True."""
torch.manual_seed(0)
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
original = self.transformers_class.from_pretrained(model_id, use_cache=False)
adapted = get_peft_model(
original, AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
)
adapted = adapted.to(self.torch_device)
expected = adapted.generate(input_ids=input_ids, max_length=8)
# Set use_cache = True and generate output again.
adapted.base_model.config.use_cache = True
actual = adapted.generate(input_ids=input_ids, max_length=8)
assert_close(expected, actual, rtol=0, atol=0)
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_bf16_inference(self, model_id):
if self.torch_device == "mps":
return pytest.skip("Skipping bf16 test on MPS")
"""Test that AdaptionPrompt works when Llama using a half-precision model."""
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
original = self.transformers_class.from_pretrained(model_id, torch_dtype=torch.bfloat16)
adapted = get_peft_model(
original, AdaptionPromptConfig(adapter_layers=2, adapter_len=4, task_type="CAUSAL_LM")
)
adapted = adapted.to(self.torch_device)
adapted.generate(input_ids=input_ids) # does not raise
@pytest.mark.xfail(reason="currently this fails because scores are zeroed out", raises=AssertionError)
@pytest.mark.parametrize("model_id", MODELS_TO_TEST)
def test_disable_adapter(self, model_id):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
dummy_input = torch.LongTensor([[1, 1, 1]]).to(self.torch_device)
output_before = model(dummy_input).logits
config = AdaptionPromptConfig(adapter_layers=1, adapter_len=4, task_type="CAUSAL_LM")
model = get_peft_model(model, config).to(self.torch_device)
output_peft = model(dummy_input).logits
# TODO currently this fails because scores are zeroed out:
# https://github.com/huggingface/peft/blob/062d95a09eb5d1de35c0e5e23d4387daba99e2db/src/peft/tuners/adaption_prompt.py#L303
# This is fine for users but makes it difficult to test if anything happens. In the future, we will have a clean
# way to control initialization. Until then, this test is expected to fail.
assert not torch.allclose(output_before, output_peft)
with model.disable_adapter():
output_peft_disabled = model(dummy_input).logits
assert torch.allclose(output_before, output_peft_disabled)
|