forked from Arnav0400/peft
-
Notifications
You must be signed in to change notification settings - Fork 0
/
testing_common.py
487 lines (380 loc) · 17.6 KB
/
testing_common.py
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
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
# coding=utf-8
# 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
from collections import OrderedDict
import torch
from peft import (
LoraConfig,
PeftModel,
PrefixTuningConfig,
PromptEncoderConfig,
PromptTuningConfig,
get_peft_model,
get_peft_model_state_dict,
prepare_model_for_int8_training,
)
CONFIG_CLASSES = (
LoraConfig,
PrefixTuningConfig,
PromptEncoderConfig,
PromptTuningConfig,
)
CONFIG_TESTING_KWARGS = (
{
"r": 8,
"lora_alpha": 32,
"target_modules": None,
"lora_dropout": 0.05,
"bias": "none",
},
{
"num_virtual_tokens": 10,
},
{
"num_virtual_tokens": 10,
"encoder_hidden_size": 32,
},
{
"num_virtual_tokens": 10,
},
)
CLASSES_MAPPING = {
"lora": (LoraConfig, CONFIG_TESTING_KWARGS[0]),
"prefix_tuning": (PrefixTuningConfig, CONFIG_TESTING_KWARGS[1]),
"prompt_encoder": (PromptEncoderConfig, CONFIG_TESTING_KWARGS[2]),
"prompt_tuning": (PromptTuningConfig, CONFIG_TESTING_KWARGS[3]),
}
# Adapted from https://github.com/huggingface/transformers/blob/48327c57182fdade7f7797d1eaad2d166de5c55b/src/transformers/activations.py#LL166C7-L166C22
class ClassInstantier(OrderedDict):
def __getitem__(self, key, *args, **kwargs):
# check if any of the kwargs is inside the config class kwargs
if any(kwarg in self[key][1] for kwarg in kwargs):
new_config_kwargs = self[key][1].copy()
new_config_kwargs.update(kwargs)
return (self[key][0], new_config_kwargs)
return super().__getitem__(key, *args, **kwargs)
def get_grid_parameters(self, grid_parameters, filter_params_func=None):
r"""
Returns a list of all possible combinations of the parameters in the config classes.
Args:
grid_parameters (`dict`):
A dictionary containing the parameters to be tested. There should be at least the key "model_ids" which
contains a list of model ids to be tested. The other keys should be the name of the config class
post-fixed with "_kwargs" and the value should be a dictionary containing the parameters to be tested
for that config class.
filter_params_func (`callable`, `optional`):
A function that takes a list of tuples and returns a list of tuples. This function is used to filter
out the tests that needs for example to be skipped.
Returns:
generated_tests (`list`):
A list of tuples containing the name of the test, the model id, the config class and the config class
kwargs.
"""
generated_tests = []
model_list = grid_parameters["model_ids"]
task_type = grid_parameters["task_type"] if "task_type" in grid_parameters else None
for model_id in model_list:
for key, value in self.items():
if "{}_kwargs".format(key) in grid_parameters:
peft_configs = []
current_peft_config = value[1].copy()
for current_key, current_value in grid_parameters[f"{key}_kwargs"].items():
for kwarg in current_value:
current_peft_config.update({current_key: kwarg})
if task_type is not None:
current_peft_config.update({"task_type": task_type})
peft_configs.append(current_peft_config.copy())
else:
current_peft_config = value[1].copy()
if task_type is not None:
current_peft_config.update({"task_type": task_type})
peft_configs = [current_peft_config]
for peft_config in peft_configs:
generated_tests.append((f"test_{model_id}_{key}", model_id, value[0], peft_config))
if filter_params_func is not None:
generated_tests = filter_params_func(generated_tests)
return generated_tests
PeftTestConfigManager = ClassInstantier(CLASSES_MAPPING)
class PeftCommonTester:
r"""
A large testing suite for testing common functionality of the PEFT models.
Attributes:
torch_device (`torch.device`):
The device on which the tests will be run.
transformers_class (`transformers.PreTrainedModel`):
The transformers class that is being tested.
"""
torch_device = "cuda" if torch.cuda.is_available() else "cpu"
transformers_class = None
def prepare_inputs_for_common(self):
raise NotImplementedError
def _test_model_attr(self, model_id, config_cls, config_kwargs):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
self.assertTrue(hasattr(model, "save_pretrained"))
self.assertTrue(hasattr(model, "from_pretrained"))
self.assertTrue(hasattr(model, "push_to_hub"))
def _test_prepare_for_training(self, model_id, config_cls, config_kwargs):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
dummy_input = self.prepare_inputs_for_testing()
dummy_output = model.get_input_embeddings()(dummy_input["input_ids"])
self.assertTrue(not dummy_output.requires_grad)
# load with `prepare_model_for_int8_training`
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
model = prepare_model_for_int8_training(model)
for param in model.parameters():
self.assertTrue(not param.requires_grad)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
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 = self.prepare_inputs_for_testing()
dummy_output = model.get_input_embeddings()(dummy_input["input_ids"])
self.assertTrue(dummy_output.requires_grad)
def _test_save_pretrained(self, model_id, config_cls, config_kwargs):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
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
self.assertEqual(state_dict.keys(), state_dict_from_pretrained.keys())
# check if tensors equal
for key in state_dict.keys():
self.assertTrue(
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
self.assertTrue(os.path.exists(os.path.join(tmp_dirname, "adapter_model.bin")))
# check if `adapter_config.json` is present
self.assertTrue(os.path.exists(os.path.join(tmp_dirname, "adapter_config.json")))
# check if `pytorch_model.bin` is not present
self.assertFalse(os.path.exists(os.path.join(tmp_dirname, "pytorch_model.bin")))
# check if `config.json` is not present
self.assertFalse(os.path.exists(os.path.join(tmp_dirname, "config.json")))
def _test_merge_layers(self, model_id, config_cls, config_kwargs):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
if config.peft_type != "LORA":
with self.assertRaises(AttributeError):
model = model.merge_and_unload()
elif model.config.model_type == "gpt2":
with self.assertRaises(ValueError):
model = model.merge_and_unload()
else:
dummy_input = self.prepare_inputs_for_testing()
model.eval()
logits_lora = model(**dummy_input)[0]
model = model.merge_and_unload()
logits_merged = model(**dummy_input)[0]
transformers_model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
logits_transformers = transformers_model(**dummy_input)[0]
self.assertTrue(torch.allclose(logits_lora, logits_merged, atol=1e-4, rtol=1e-4))
self.assertFalse(torch.allclose(logits_merged, logits_transformers, atol=1e-10, rtol=1e-10))
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(tmp_dirname).to(self.torch_device)
logits_merged_from_pretrained = model_from_pretrained(**dummy_input)[0]
self.assertTrue(torch.allclose(logits_merged, logits_merged_from_pretrained, atol=1e-4, rtol=1e-4))
def _test_generate(self, model_id, config_cls, config_kwargs):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `generate` works
_ = model.generate(**inputs)
with self.assertRaises(TypeError):
# check if `generate` raises an error if no positional arguments are passed
_ = model.generate(inputs["input_ids"])
def _test_generate_half_prec(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig, PrefixTuningConfig):
return
model = self.transformers_class.from_pretrained(model_id, torch_dtype=torch.bfloat16)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
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)
with self.assertRaises(TypeError):
# check if `generate` raises an error if no positional arguments are passed
_ = model.generate(input_ids, attention_mask=attention_mask)
def _test_training(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
loss = output.sum()
loss.backward()
for n, param in model.named_parameters():
if "lora" in n:
self.assertIsNotNone(param.grad)
else:
self.assertIsNone(param.grad)
def _test_inference_safetensors(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return
config = config_cls(
base_model_name_or_path=model_id,
layers_to_transform=[0],
**config_kwargs,
)
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
logits = output[0]
loss = output.sum()
loss.backward()
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname, safe_serialization=True)
self.assertTrue("adapter_model.safetensors" in os.listdir(tmp_dirname))
self.assertTrue("adapter_model.bin" not in os.listdir(tmp_dirname))
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname).to(self.torch_device)
logits_from_pretrained = model_from_pretrained(**inputs)[0][0]
self.assertTrue(torch.allclose(logits, logits_from_pretrained, atol=1e-4, rtol=1e-4))
def _test_training_layer_indexing(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return
config = config_cls(
base_model_name_or_path=model_id,
layers_to_transform=[0],
**config_kwargs,
)
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
logits = output[0]
loss = output.sum()
loss.backward()
nb_trainable = 0
for n, param in model.named_parameters():
if "lora" in n:
self.assertIsNotNone(param.grad)
nb_trainable += 1
else:
self.assertIsNone(param.grad)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname).to(self.torch_device)
logits_from_pretrained = model_from_pretrained(**inputs)[0][0]
self.assertTrue(torch.allclose(logits, logits_from_pretrained, atol=1e-4, rtol=1e-4))
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
nb_trainable_all = 0
for n, param in model.named_parameters():
if "lora" in n:
nb_trainable_all += 1
self.assertLess(nb_trainable, nb_trainable_all)
def _test_training_gradient_checkpointing(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return
model = self.transformers_class.from_pretrained(model_id)
if not model.supports_gradient_checkpointing:
return
model.gradient_checkpointing_enable()
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
loss = output.sum()
loss.backward()
for n, param in model.named_parameters():
if "lora" in n:
self.assertIsNotNone(param.grad)
else:
self.assertIsNone(param.grad)
def _test_peft_model_device_map(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
_ = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname, device_map={"": "cpu"}).to(
self.torch_device
)