"docs/source/en/using-diffusers/kerascv.mdx" did not exist on "dd3cae33279076d813000f03c63e4ee84ce0fa77"
utils.py 78 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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.
Sayak Paul's avatar
Sayak Paul committed
15
import inspect
16
17
18
import os
import tempfile
import unittest
UmerHA's avatar
UmerHA committed
19
from itertools import product
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

import numpy as np
import torch

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    LCMScheduler,
    UNet2DConditionModel,
)
from diffusers.utils.import_utils import is_peft_available
from diffusers.utils.testing_utils import (
    floats_tensor,
    require_peft_backend,
    require_peft_version_greater,
35
    skip_mps,
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
    torch_device,
)


if is_peft_available():
    from peft import LoraConfig
    from peft.tuners.tuners_utils import BaseTunerLayer
    from peft.utils import get_peft_model_state_dict


def state_dicts_almost_equal(sd1, sd2):
    sd1 = dict(sorted(sd1.items()))
    sd2 = dict(sorted(sd2.items()))

    models_are_equal = True
    for ten1, ten2 in zip(sd1.values(), sd2.values()):
        if (ten1 - ten2).abs().max() > 1e-3:
            models_are_equal = False

    return models_are_equal


def check_if_lora_correctly_set(model) -> bool:
    """
    Checks if the LoRA layers are correctly set with peft
    """
    for module in model.modules():
        if isinstance(module, BaseTunerLayer):
            return True
    return False


@require_peft_backend
class PeftLoraLoaderMixinTests:
    pipeline_class = None
Aryan's avatar
Aryan committed
71

72
73
    scheduler_cls = None
    scheduler_kwargs = None
Aryan's avatar
Aryan committed
74
    scheduler_classes = [DDIMScheduler, LCMScheduler]
Sayak Paul's avatar
Sayak Paul committed
75

76
    has_two_text_encoders = False
77
    has_three_text_encoders = False
Sayak Paul's avatar
Sayak Paul committed
78
79
80
81
82
83
84
    text_encoder_cls, text_encoder_id = None, None
    text_encoder_2_cls, text_encoder_2_id = None, None
    text_encoder_3_cls, text_encoder_3_id = None, None
    tokenizer_cls, tokenizer_id = None, None
    tokenizer_2_cls, tokenizer_2_id = None, None
    tokenizer_3_cls, tokenizer_3_id = None, None

85
    unet_kwargs = None
Sayak Paul's avatar
Sayak Paul committed
86
    transformer_cls = None
87
    transformer_kwargs = None
Aryan's avatar
Aryan committed
88
    vae_cls = AutoencoderKL
89
90
    vae_kwargs = None

Aryan's avatar
Aryan committed
91
92
    text_encoder_target_modules = ["q_proj", "k_proj", "v_proj", "out_proj"]

93
    def get_dummy_components(self, scheduler_cls=None, use_dora=False):
94
95
96
97
98
        if self.unet_kwargs and self.transformer_kwargs:
            raise ValueError("Both `unet_kwargs` and `transformer_kwargs` cannot be specified.")
        if self.has_two_text_encoders and self.has_three_text_encoders:
            raise ValueError("Both `has_two_text_encoders` and `has_three_text_encoders` cannot be True.")

99
        scheduler_cls = self.scheduler_cls if scheduler_cls is None else scheduler_cls
100
101
102
        rank = 4

        torch.manual_seed(0)
103
104
105
        if self.unet_kwargs is not None:
            unet = UNet2DConditionModel(**self.unet_kwargs)
        else:
Sayak Paul's avatar
Sayak Paul committed
106
            transformer = self.transformer_cls(**self.transformer_kwargs)
107
108
109
110

        scheduler = scheduler_cls(**self.scheduler_kwargs)

        torch.manual_seed(0)
Aryan's avatar
Aryan committed
111
        vae = self.vae_cls(**self.vae_kwargs)
112

Sayak Paul's avatar
Sayak Paul committed
113
114
        text_encoder = self.text_encoder_cls.from_pretrained(self.text_encoder_id)
        tokenizer = self.tokenizer_cls.from_pretrained(self.tokenizer_id)
115

Sayak Paul's avatar
Sayak Paul committed
116
117
118
        if self.text_encoder_2_cls is not None:
            text_encoder_2 = self.text_encoder_2_cls.from_pretrained(self.text_encoder_2_id)
            tokenizer_2 = self.tokenizer_2_cls.from_pretrained(self.tokenizer_2_id)
119

Sayak Paul's avatar
Sayak Paul committed
120
121
122
        if self.text_encoder_3_cls is not None:
            text_encoder_3 = self.text_encoder_3_cls.from_pretrained(self.text_encoder_3_id)
            tokenizer_3 = self.tokenizer_3_cls.from_pretrained(self.tokenizer_3_id)
123

124
125
126
        text_lora_config = LoraConfig(
            r=rank,
            lora_alpha=rank,
Aryan's avatar
Aryan committed
127
            target_modules=self.text_encoder_target_modules,
128
            init_lora_weights=False,
129
            use_dora=use_dora,
130
131
        )

132
        denoiser_lora_config = LoraConfig(
133
134
135
136
137
            r=rank,
            lora_alpha=rank,
            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
            init_lora_weights=False,
            use_dora=use_dora,
138
139
        )

Sayak Paul's avatar
Sayak Paul committed
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
        pipeline_components = {
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
        }
        # Denoiser
        if self.unet_kwargs is not None:
            pipeline_components.update({"unet": unet})
        elif self.transformer_kwargs is not None:
            pipeline_components.update({"transformer": transformer})

        # Remaining text encoders.
        if self.text_encoder_2_cls is not None:
            pipeline_components.update({"tokenizer_2": tokenizer_2, "text_encoder_2": text_encoder_2})
        if self.text_encoder_3_cls is not None:
            pipeline_components.update({"tokenizer_3": tokenizer_3, "text_encoder_3": text_encoder_3})

        # Remaining stuff
        init_params = inspect.signature(self.pipeline_class.__init__).parameters
        if "safety_checker" in init_params:
            pipeline_components.update({"safety_checker": None})
        if "feature_extractor" in init_params:
            pipeline_components.update({"feature_extractor": None})
        if "image_encoder" in init_params:
            pipeline_components.update({"image_encoder": None})
166

167
        return pipeline_components, text_lora_config, denoiser_lora_config
168

Sayak Paul's avatar
Sayak Paul committed
169
170
171
172
    @property
    def output_shape(self):
        raise NotImplementedError

173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
    def get_dummy_inputs(self, with_generator=True):
        batch_size = 1
        sequence_length = 10
        num_channels = 4
        sizes = (32, 32)

        generator = torch.manual_seed(0)
        noise = floats_tensor((batch_size, num_channels) + sizes)
        input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)

        pipeline_inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "num_inference_steps": 5,
            "guidance_scale": 6.0,
            "output_type": "np",
        }
        if with_generator:
            pipeline_inputs.update({"generator": generator})

        return noise, input_ids, pipeline_inputs

194
    # Copied from: https://colab.research.google.com/gist/sayakpaul/df2ef6e1ae6d8c10a49d859883b10860/scratchpad.ipynb
195
196
197
198
199
200
201
202
203
204
205
206
207
    def get_dummy_tokens(self):
        max_seq_length = 77

        inputs = torch.randint(2, 56, size=(1, max_seq_length), generator=torch.manual_seed(0))

        prepared_inputs = {}
        prepared_inputs["input_ids"] = inputs
        return prepared_inputs

    def test_simple_inference(self):
        """
        Tests a simple inference and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
208
        for scheduler_cls in self.scheduler_classes:
209
210
211
212
213
214
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)

            _, _, inputs = self.get_dummy_inputs()
Aryan's avatar
Aryan committed
215
            output_no_lora = pipe(**inputs)[0]
Sayak Paul's avatar
Sayak Paul committed
216
            self.assertTrue(output_no_lora.shape == self.output_shape)
217
218
219
220
221
222

    def test_simple_inference_with_text_lora(self):
        """
        Tests a simple inference with lora attached on the text encoder
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
223
        for scheduler_cls in self.scheduler_classes:
224
225
226
227
228
229
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
230
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
231
            self.assertTrue(output_no_lora.shape == self.output_shape)
232
233
234
235

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

236
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
237
238
239
240
241
242
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
243

Aryan's avatar
Aryan committed
244
            output_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
245
246
247
248
249
250
251
252
253
            self.assertTrue(
                not np.allclose(output_lora, output_no_lora, atol=1e-3, rtol=1e-3), "Lora should change the output"
            )

    def test_simple_inference_with_text_lora_and_scale(self):
        """
        Tests a simple inference with lora attached on the text encoder + scale argument
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
254
        call_signature_keys = inspect.signature(self.pipeline_class.__call__).parameters.keys()
Aryan's avatar
Aryan committed
255
256

        # TODO(diffusers): Discuss a common naming convention across library for 1.0.0 release
Aryan's avatar
Aryan committed
257
258
259
260
261
262
        for possible_attention_kwargs in ["cross_attention_kwargs", "joint_attention_kwargs", "attention_kwargs"]:
            if possible_attention_kwargs in call_signature_keys:
                attention_kwargs_name = possible_attention_kwargs
                break
        assert attention_kwargs_name is not None

Aryan's avatar
Aryan committed
263
        for scheduler_cls in self.scheduler_classes:
264
265
266
267
268
269
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
270
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
271
            self.assertTrue(output_no_lora.shape == self.output_shape)
272
273
274
275

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

276
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
277
278
279
280
281
282
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
283

Aryan's avatar
Aryan committed
284
            output_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
285
286
287
288
            self.assertTrue(
                not np.allclose(output_lora, output_no_lora, atol=1e-3, rtol=1e-3), "Lora should change the output"
            )

Aryan's avatar
Aryan committed
289
290
291
            attention_kwargs = {attention_kwargs_name: {"scale": 0.5}}
            output_lora_scale = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]

292
293
294
295
296
            self.assertTrue(
                not np.allclose(output_lora, output_lora_scale, atol=1e-3, rtol=1e-3),
                "Lora + scale should change the output",
            )

Aryan's avatar
Aryan committed
297
298
299
            attention_kwargs = {attention_kwargs_name: {"scale": 0.0}}
            output_lora_0_scale = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]

300
301
302
303
304
305
306
307
308
309
            self.assertTrue(
                np.allclose(output_no_lora, output_lora_0_scale, atol=1e-3, rtol=1e-3),
                "Lora + 0 scale should lead to same result as no LoRA",
            )

    def test_simple_inference_with_text_lora_fused(self):
        """
        Tests a simple inference with lora attached into text encoder + fuses the lora weights into base model
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
310
        for scheduler_cls in self.scheduler_classes:
311
312
313
314
315
316
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
317
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
318
            self.assertTrue(output_no_lora.shape == self.output_shape)
319
320
321
322

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

323
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
324
325
326
327
328
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
329
330
331
332
333

            pipe.fuse_lora()
            # Fusing should still keep the LoRA layers
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

334
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
335
336
337
338
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
339

Aryan's avatar
Aryan committed
340
            ouput_fused = pipe(**inputs, generator=torch.manual_seed(0))[0]
341
342
343
344
345
346
347
348
349
            self.assertFalse(
                np.allclose(ouput_fused, output_no_lora, atol=1e-3, rtol=1e-3), "Fused lora should change the output"
            )

    def test_simple_inference_with_text_lora_unloaded(self):
        """
        Tests a simple inference with lora attached to text encoder, then unloads the lora weights
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
350
        for scheduler_cls in self.scheduler_classes:
351
352
353
354
355
356
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
357
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
358
            self.assertTrue(output_no_lora.shape == self.output_shape)
359

Aryan's avatar
Aryan committed
360
361
362
363
364
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
365

366
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
367
368
369
370
371
372
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
373
374
375
376
377
378
379

            pipe.unload_lora_weights()
            # unloading should remove the LoRA layers
            self.assertFalse(
                check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly unloaded in text encoder"
            )

380
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
381
382
383
384
385
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertFalse(
                        check_if_lora_correctly_set(pipe.text_encoder_2),
                        "Lora not correctly unloaded in text encoder 2",
                    )
386

Aryan's avatar
Aryan committed
387
            ouput_unloaded = pipe(**inputs, generator=torch.manual_seed(0))[0]
388
389
390
391
392
393
394
395
396
            self.assertTrue(
                np.allclose(ouput_unloaded, output_no_lora, atol=1e-3, rtol=1e-3),
                "Fused lora should change the output",
            )

    def test_simple_inference_with_text_lora_save_load(self):
        """
        Tests a simple usecase where users could use saving utilities for LoRA.
        """
Aryan's avatar
Aryan committed
397
        for scheduler_cls in self.scheduler_classes:
398
399
400
401
402
403
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
404
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
405
            self.assertTrue(output_no_lora.shape == self.output_shape)
406

Aryan's avatar
Aryan committed
407
408
409
410
411
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
412

413
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
414
415
416
417
418
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
419

Aryan's avatar
Aryan committed
420
            images_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
421
422
423

            with tempfile.TemporaryDirectory() as tmpdirname:
                text_encoder_state_dict = get_peft_model_state_dict(pipe.text_encoder)
424
                if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
425
426
                    if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                        text_encoder_2_state_dict = get_peft_model_state_dict(pipe.text_encoder_2)
427

Sayak Paul's avatar
Sayak Paul committed
428
429
430
431
432
433
                        self.pipeline_class.save_lora_weights(
                            save_directory=tmpdirname,
                            text_encoder_lora_layers=text_encoder_state_dict,
                            text_encoder_2_lora_layers=text_encoder_2_state_dict,
                            safe_serialization=False,
                        )
434
435
436
437
438
439
440
                else:
                    self.pipeline_class.save_lora_weights(
                        save_directory=tmpdirname,
                        text_encoder_lora_layers=text_encoder_state_dict,
                        safe_serialization=False,
                    )

Sayak Paul's avatar
Sayak Paul committed
441
442
443
444
445
446
447
448
                if self.has_two_text_encoders:
                    if "text_encoder_2" not in self.pipeline_class._lora_loadable_modules:
                        self.pipeline_class.save_lora_weights(
                            save_directory=tmpdirname,
                            text_encoder_lora_layers=text_encoder_state_dict,
                            safe_serialization=False,
                        )

449
450
451
452
453
                self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.bin")))
                pipe.unload_lora_weights()

                pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.bin"))

Aryan's avatar
Aryan committed
454
            images_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0))[0]
455
456
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

457
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
458
459
460
461
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
462
463
464
465
466
467

            self.assertTrue(
                np.allclose(images_lora, images_lora_from_pretrained, atol=1e-3, rtol=1e-3),
                "Loading from saved checkpoints should give same results.",
            )

468
469
470
471
472
473
    def test_simple_inference_with_partial_text_lora(self):
        """
        Tests a simple inference with lora attached on the text encoder
        with different ranks and some adapters removed
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
474
        for scheduler_cls in self.scheduler_classes:
475
            components, _, _ = self.get_dummy_components(scheduler_cls)
476
            # Verify `StableDiffusionLoraLoaderMixin.load_lora_into_text_encoder` handles different ranks per module (PR#8324).
477
478
479
480
481
482
483
484
485
486
487
488
489
            text_lora_config = LoraConfig(
                r=4,
                rank_pattern={"q_proj": 1, "k_proj": 2, "v_proj": 3},
                lora_alpha=4,
                target_modules=["q_proj", "k_proj", "v_proj", "out_proj"],
                init_lora_weights=False,
                use_dora=False,
            )
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
490
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
491
            self.assertTrue(output_no_lora.shape == self.output_shape)
492
493
494
495
496
497
498
499
500
501
502

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
            # Gather the state dict for the PEFT model, excluding `layers.4`, to ensure `load_lora_into_text_encoder`
            # supports missing layers (PR#8324).
            state_dict = {
                f"text_encoder.{module_name}": param
                for module_name, param in get_peft_model_state_dict(pipe.text_encoder).items()
                if "text_model.encoder.layers.4" not in module_name
            }

503
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
504
505
506
507
508
509
510
511
512
513
514
515
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
                    state_dict.update(
                        {
                            f"text_encoder_2.{module_name}": param
                            for module_name, param in get_peft_model_state_dict(pipe.text_encoder_2).items()
                            if "text_model.encoder.layers.4" not in module_name
                        }
                    )
516

Aryan's avatar
Aryan committed
517
            output_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
518
519
520
521
522
523
524
525
            self.assertTrue(
                not np.allclose(output_lora, output_no_lora, atol=1e-3, rtol=1e-3), "Lora should change the output"
            )

            # Unload lora and load it back using the pipe.load_lora_weights machinery
            pipe.unload_lora_weights()
            pipe.load_lora_weights(state_dict)

Aryan's avatar
Aryan committed
526
            output_partial_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
527
528
529
530
531
            self.assertTrue(
                not np.allclose(output_partial_lora, output_lora, atol=1e-3, rtol=1e-3),
                "Removing adapters should change the output",
            )

532
533
534
535
    def test_simple_inference_save_pretrained(self):
        """
        Tests a simple usecase where users could use saving utilities for LoRA through save_pretrained
        """
Aryan's avatar
Aryan committed
536
        for scheduler_cls in self.scheduler_classes:
537
538
539
540
541
542
            components, text_lora_config, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
543
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
544
            self.assertTrue(output_no_lora.shape == self.output_shape)
545
546
547
548

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")

549
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
550
551
552
553
554
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
555

Aryan's avatar
Aryan committed
556
            images_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
557
558
559
560
561
562
563
564
565
566
567
568

            with tempfile.TemporaryDirectory() as tmpdirname:
                pipe.save_pretrained(tmpdirname)

                pipe_from_pretrained = self.pipeline_class.from_pretrained(tmpdirname)
                pipe_from_pretrained.to(torch_device)

            self.assertTrue(
                check_if_lora_correctly_set(pipe_from_pretrained.text_encoder),
                "Lora not correctly set in text encoder",
            )

569
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
570
571
572
573
574
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe_from_pretrained.text_encoder_2),
                        "Lora not correctly set in text encoder 2",
                    )
575

Aryan's avatar
Aryan committed
576
            images_lora_save_pretrained = pipe_from_pretrained(**inputs, generator=torch.manual_seed(0))[0]
577
578
579
580
581
582

            self.assertTrue(
                np.allclose(images_lora, images_lora_save_pretrained, atol=1e-3, rtol=1e-3),
                "Loading from saved checkpoints should give same results.",
            )

583
    def test_simple_inference_with_text_denoiser_lora_save_load(self):
584
585
586
        """
        Tests a simple usecase where users could use saving utilities for LoRA for Unet + text encoder
        """
Aryan's avatar
Aryan committed
587
        for scheduler_cls in self.scheduler_classes:
588
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
589
590
591
592
593
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
594
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
595
            self.assertTrue(output_no_lora.shape == self.output_shape)
596

Aryan's avatar
Aryan committed
597
598
599
600
601
602
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
603
604
605
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
606

607
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
608
609
610
611
612
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
613

Aryan's avatar
Aryan committed
614
            images_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
615
616

            with tempfile.TemporaryDirectory() as tmpdirname:
Aryan's avatar
Aryan committed
617
618
619
620
621
                text_encoder_state_dict = (
                    get_peft_model_state_dict(pipe.text_encoder)
                    if "text_encoder" in self.pipeline_class._lora_loadable_modules
                    else None
                )
622

Aryan's avatar
Aryan committed
623
                denoiser_state_dict = get_peft_model_state_dict(denoiser)
624

Sayak Paul's avatar
Sayak Paul committed
625
626
627
628
                saving_kwargs = {
                    "save_directory": tmpdirname,
                    "safe_serialization": False,
                }
629

Aryan's avatar
Aryan committed
630
631
632
                if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                    saving_kwargs.update({"text_encoder_lora_layers": text_encoder_state_dict})

Sayak Paul's avatar
Sayak Paul committed
633
634
                if self.unet_kwargs is not None:
                    saving_kwargs.update({"unet_lora_layers": denoiser_state_dict})
635
                else:
Sayak Paul's avatar
Sayak Paul committed
636
637
638
639
640
641
642
643
                    saving_kwargs.update({"transformer_lora_layers": denoiser_state_dict})

                if self.has_two_text_encoders or self.has_three_text_encoders:
                    if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                        text_encoder_2_state_dict = get_peft_model_state_dict(pipe.text_encoder_2)
                        saving_kwargs.update({"text_encoder_2_lora_layers": text_encoder_2_state_dict})

                self.pipeline_class.save_lora_weights(**saving_kwargs)
644
645
646
647
648
649

                self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.bin")))
                pipe.unload_lora_weights()

                pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.bin"))

Aryan's avatar
Aryan committed
650
651
652
653
654
655
656
            images_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0))[0]

            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
657
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser")
658

659
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
660
661
662
663
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
664
665
666
667
668
669

            self.assertTrue(
                np.allclose(images_lora, images_lora_from_pretrained, atol=1e-3, rtol=1e-3),
                "Loading from saved checkpoints should give same results.",
            )

670
    def test_simple_inference_with_text_denoiser_lora_and_scale(self):
671
672
673
674
        """
        Tests a simple inference with lora attached on the text encoder + Unet + scale argument
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
675
676
677
678
679
680
681
        call_signature_keys = inspect.signature(self.pipeline_class.__call__).parameters.keys()
        for possible_attention_kwargs in ["cross_attention_kwargs", "joint_attention_kwargs", "attention_kwargs"]:
            if possible_attention_kwargs in call_signature_keys:
                attention_kwargs_name = possible_attention_kwargs
                break
        assert attention_kwargs_name is not None

Aryan's avatar
Aryan committed
682
        for scheduler_cls in self.scheduler_classes:
683
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
684
685
686
687
688
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
689
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
690
            self.assertTrue(output_no_lora.shape == self.output_shape)
691

Aryan's avatar
Aryan committed
692
693
694
695
696
697
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
698
699
700
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
701

702
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
703
704
705
706
707
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
708

Aryan's avatar
Aryan committed
709
            output_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
710
711
712
713
            self.assertTrue(
                not np.allclose(output_lora, output_no_lora, atol=1e-3, rtol=1e-3), "Lora should change the output"
            )

Aryan's avatar
Aryan committed
714
715
716
            attention_kwargs = {attention_kwargs_name: {"scale": 0.5}}
            output_lora_scale = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]

717
718
719
720
721
            self.assertTrue(
                not np.allclose(output_lora, output_lora_scale, atol=1e-3, rtol=1e-3),
                "Lora + scale should change the output",
            )

Aryan's avatar
Aryan committed
722
723
724
            attention_kwargs = {attention_kwargs_name: {"scale": 0.0}}
            output_lora_0_scale = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]

725
726
727
728
729
            self.assertTrue(
                np.allclose(output_no_lora, output_lora_0_scale, atol=1e-3, rtol=1e-3),
                "Lora + 0 scale should lead to same result as no LoRA",
            )

Aryan's avatar
Aryan committed
730
731
732
733
734
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                self.assertTrue(
                    pipe.text_encoder.text_model.encoder.layers[0].self_attn.q_proj.scaling["default"] == 1.0,
                    "The scaling parameter has not been correctly restored!",
                )
735

736
    def test_simple_inference_with_text_lora_denoiser_fused(self):
737
738
739
740
        """
        Tests a simple inference with lora attached into text encoder + fuses the lora weights into base model
        and makes sure it works as expected - with unet
        """
Aryan's avatar
Aryan committed
741
        for scheduler_cls in self.scheduler_classes:
742
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
743
744
745
746
747
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
748
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
749
            self.assertTrue(output_no_lora.shape == self.output_shape)
750

Aryan's avatar
Aryan committed
751
752
753
754
755
756
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
757
758
759
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
760

761
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
762
763
764
765
766
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
767

Aryan's avatar
Aryan committed
768
769
            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules)

770
            # Fusing should still keep the LoRA layers
Aryan's avatar
Aryan committed
771
772
773
774
775
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
776
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser")
777

778
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
779
780
781
782
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
783

Aryan's avatar
Aryan committed
784
            output_fused = pipe(**inputs, generator=torch.manual_seed(0))[0]
785
            self.assertFalse(
Aryan's avatar
Aryan committed
786
                np.allclose(output_fused, output_no_lora, atol=1e-3, rtol=1e-3), "Fused lora should change the output"
787
788
            )

789
    def test_simple_inference_with_text_denoiser_lora_unloaded(self):
790
791
792
793
        """
        Tests a simple inference with lora attached to text encoder and unet, then unloads the lora weights
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
794
        for scheduler_cls in self.scheduler_classes:
795
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
796
797
798
799
800
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
801
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
802
            self.assertTrue(output_no_lora.shape == self.output_shape)
803

Aryan's avatar
Aryan committed
804
805
806
807
808
809
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
810
811
812
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
813

814
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
815
816
817
818
819
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
820
821
822
823
824
825

            pipe.unload_lora_weights()
            # unloading should remove the LoRA layers
            self.assertFalse(
                check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly unloaded in text encoder"
            )
Aryan's avatar
Aryan committed
826
            self.assertFalse(check_if_lora_correctly_set(denoiser), "Lora not correctly unloaded in denoiser")
827

828
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
829
830
831
832
833
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertFalse(
                        check_if_lora_correctly_set(pipe.text_encoder_2),
                        "Lora not correctly unloaded in text encoder 2",
                    )
834

Aryan's avatar
Aryan committed
835
            output_unloaded = pipe(**inputs, generator=torch.manual_seed(0))[0]
836
            self.assertTrue(
Aryan's avatar
Aryan committed
837
                np.allclose(output_unloaded, output_no_lora, atol=1e-3, rtol=1e-3),
838
839
840
                "Fused lora should change the output",
            )

Aryan's avatar
Aryan committed
841
842
843
    def test_simple_inference_with_text_denoiser_lora_unfused(
        self, expected_atol: float = 1e-3, expected_rtol: float = 1e-3
    ):
844
845
846
847
        """
        Tests a simple inference with lora attached to text encoder and unet, then unloads the lora weights
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
848
        for scheduler_cls in self.scheduler_classes:
849
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
850
851
852
853
854
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
855
856
857
858
859
860
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
861
862
863
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
864

865
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
866
867
868
869
870
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
871

Aryan's avatar
Aryan committed
872
873
            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules)
            output_fused_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
874

Aryan's avatar
Aryan committed
875
876
            pipe.unfuse_lora(components=self.pipeline_class._lora_loadable_modules)
            output_unfused_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
877
878

            # unloading should remove the LoRA layers
Aryan's avatar
Aryan committed
879
880
881
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Unfuse should still keep LoRA layers")

Aryan's avatar
Aryan committed
882
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Unfuse should still keep LoRA layers")
883

884
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
885
886
887
888
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Unfuse should still keep LoRA layers"
                    )
889
890
891

            # Fuse and unfuse should lead to the same results
            self.assertTrue(
Aryan's avatar
Aryan committed
892
893
                np.allclose(output_fused_lora, output_unfused_lora, atol=expected_atol, rtol=expected_rtol),
                "Fused lora should not change the output",
894
895
            )

896
    def test_simple_inference_with_text_denoiser_multi_adapter(self):
897
898
899
900
        """
        Tests a simple inference with lora attached to text encoder and unet, attaches
        multiple adapters and set them
        """
Aryan's avatar
Aryan committed
901
        for scheduler_cls in self.scheduler_classes:
902
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
903
904
905
906
907
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
908
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
909

Aryan's avatar
Aryan committed
910
911
912
913
914
915
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
916

Aryan's avatar
Aryan committed
917
918
919
920
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
921

922
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
923
924
925
926
927
928
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
929
930

            pipe.set_adapters("adapter-1")
Aryan's avatar
Aryan committed
931
            output_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
932
933

            pipe.set_adapters("adapter-2")
Aryan's avatar
Aryan committed
934
            output_adapter_2 = pipe(**inputs, generator=torch.manual_seed(0))[0]
935
936

            pipe.set_adapters(["adapter-1", "adapter-2"])
Aryan's avatar
Aryan committed
937
            output_adapter_mixed = pipe(**inputs, generator=torch.manual_seed(0))[0]
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955

            # Fuse and unfuse should lead to the same results
            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
                "Adapter 1 and 2 should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 1 and mixed adapters should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_2, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 2 and mixed adapters should give different results",
            )

            pipe.disable_lora()
Aryan's avatar
Aryan committed
956
            output_disabled = pipe(**inputs, generator=torch.manual_seed(0))[0]
957
958
959
960
961
962

            self.assertTrue(
                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

963
    def test_simple_inference_with_text_denoiser_block_scale(self):
UmerHA's avatar
UmerHA committed
964
965
        """
        Tests a simple inference with lora attached to text encoder and unet, attaches
Aryan's avatar
Aryan committed
966
        one adapter and set different weights for different blocks (i.e. block lora)
UmerHA's avatar
UmerHA committed
967
        """
Aryan's avatar
Aryan committed
968
        for scheduler_cls in self.scheduler_classes:
969
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
UmerHA's avatar
UmerHA committed
970
971
972
973
974
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
975
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
976
977
978

            pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
Aryan's avatar
Aryan committed
979
980
981
982

            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
UmerHA's avatar
UmerHA committed
983

984
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
985
986
987
988
989
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
UmerHA's avatar
UmerHA committed
990
991
992

            weights_1 = {"text_encoder": 2, "unet": {"down": 5}}
            pipe.set_adapters("adapter-1", weights_1)
Aryan's avatar
Aryan committed
993
            output_weights_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
994
995
996

            weights_2 = {"unet": {"up": 5}}
            pipe.set_adapters("adapter-1", weights_2)
Aryan's avatar
Aryan committed
997
            output_weights_2 = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012

            self.assertFalse(
                np.allclose(output_weights_1, output_weights_2, atol=1e-3, rtol=1e-3),
                "LoRA weights 1 and 2 should give different results",
            )
            self.assertFalse(
                np.allclose(output_no_lora, output_weights_1, atol=1e-3, rtol=1e-3),
                "No adapter and LoRA weights 1 should give different results",
            )
            self.assertFalse(
                np.allclose(output_no_lora, output_weights_2, atol=1e-3, rtol=1e-3),
                "No adapter and LoRA weights 2 should give different results",
            )

            pipe.disable_lora()
Aryan's avatar
Aryan committed
1013
            output_disabled = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1014
1015
1016
1017
1018
1019

            self.assertTrue(
                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

1020
    def test_simple_inference_with_text_denoiser_multi_adapter_block_lora(self):
UmerHA's avatar
UmerHA committed
1021
1022
1023
1024
        """
        Tests a simple inference with lora attached to text encoder and unet, attaches
        multiple adapters and set differnt weights for different blocks (i.e. block lora)
        """
Aryan's avatar
Aryan committed
1025
        for scheduler_cls in self.scheduler_classes:
1026
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
UmerHA's avatar
UmerHA committed
1027
1028
1029
1030
1031
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1032
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1033

Aryan's avatar
Aryan committed
1034
1035
1036
1037
1038
1039
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
UmerHA's avatar
UmerHA committed
1040

Aryan's avatar
Aryan committed
1041
1042
1043
1044
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
UmerHA's avatar
UmerHA committed
1045

1046
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1047
1048
1049
1050
1051
1052
                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
UmerHA's avatar
UmerHA committed
1053
1054
1055
1056

            scales_1 = {"text_encoder": 2, "unet": {"down": 5}}
            scales_2 = {"unet": {"down": 5, "mid": 5}}

Aryan's avatar
Aryan committed
1057
1058
            pipe.set_adapters("adapter-1", scales_1)
            output_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1059
1060

            pipe.set_adapters("adapter-2", scales_2)
Aryan's avatar
Aryan committed
1061
            output_adapter_2 = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1062
1063

            pipe.set_adapters(["adapter-1", "adapter-2"], [scales_1, scales_2])
Aryan's avatar
Aryan committed
1064
            output_adapter_mixed = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082

            # Fuse and unfuse should lead to the same results
            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
                "Adapter 1 and 2 should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 1 and mixed adapters should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_2, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 2 and mixed adapters should give different results",
            )

            pipe.disable_lora()
Aryan's avatar
Aryan committed
1083
            output_disabled = pipe(**inputs, generator=torch.manual_seed(0))[0]
UmerHA's avatar
UmerHA committed
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093

            self.assertTrue(
                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

            # a mismatching number of adapter_names and adapter_weights should raise an error
            with self.assertRaises(ValueError):
                pipe.set_adapters(["adapter-1", "adapter-2"], [scales_1])

1094
    def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
UmerHA's avatar
UmerHA committed
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
        """Tests that any valid combination of lora block scales can be used in pipe.set_adapter"""

        def updown_options(blocks_with_tf, layers_per_block, value):
            """
            Generate every possible combination for how a lora weight dict for the up/down part can be.
            E.g. 2, {"block_1": 2}, {"block_1": [2,2,2]}, {"block_1": 2, "block_2": [2,2,2]}, ...
            """
            num_val = value
            list_val = [value] * layers_per_block

            node_opts = [None, num_val, list_val]
            node_opts_foreach_block = [node_opts] * len(blocks_with_tf)

            updown_opts = [num_val]
            for nodes in product(*node_opts_foreach_block):
                if all(n is None for n in nodes):
                    continue
                opt = {}
                for b, n in zip(blocks_with_tf, nodes):
                    if n is not None:
                        opt["block_" + str(b)] = n
                updown_opts.append(opt)
            return updown_opts

        def all_possible_dict_opts(unet, value):
            """
            Generate every possible combination for how a lora weight dict can be.
            E.g. 2, {"unet: {"down": 2}}, {"unet: {"down": [2,2,2]}}, {"unet: {"mid": 2, "up": [2,2,2]}}, ...
            """

            down_blocks_with_tf = [i for i, d in enumerate(unet.down_blocks) if hasattr(d, "attentions")]
            up_blocks_with_tf = [i for i, u in enumerate(unet.up_blocks) if hasattr(u, "attentions")]

            layers_per_block = unet.config.layers_per_block

            text_encoder_opts = [None, value]
            text_encoder_2_opts = [None, value]
            mid_opts = [None, value]
            down_opts = [None] + updown_options(down_blocks_with_tf, layers_per_block, value)
            up_opts = [None] + updown_options(up_blocks_with_tf, layers_per_block + 1, value)

            opts = []

            for t1, t2, d, m, u in product(text_encoder_opts, text_encoder_2_opts, down_opts, mid_opts, up_opts):
                if all(o is None for o in (t1, t2, d, m, u)):
                    continue
                opt = {}
                if t1 is not None:
                    opt["text_encoder"] = t1
                if t2 is not None:
                    opt["text_encoder_2"] = t2
                if all(o is None for o in (d, m, u)):
                    # no unet scaling
                    continue
                opt["unet"] = {}
                if d is not None:
                    opt["unet"]["down"] = d
                if m is not None:
                    opt["unet"]["mid"] = m
                if u is not None:
                    opt["unet"]["up"] = u
                opts.append(opt)

            return opts

1160
        components, text_lora_config, denoiser_lora_config = self.get_dummy_components(self.scheduler_cls)
UmerHA's avatar
UmerHA committed
1161
1162
1163
1164
1165
1166
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        _, _, inputs = self.get_dummy_inputs(with_generator=False)

        pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
Aryan's avatar
Aryan committed
1167
1168
1169

        denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
        denoiser.add_adapter(denoiser_lora_config, "adapter-1")
UmerHA's avatar
UmerHA committed
1170

1171
        if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1172
1173
1174
            lora_loadable_components = self.pipeline_class._lora_loadable_modules
            if "text_encoder_2" in lora_loadable_components:
                pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
UmerHA's avatar
UmerHA committed
1175
1176
1177
1178
1179
1180
1181
1182

        for scale_dict in all_possible_dict_opts(pipe.unet, value=1234):
            # test if lora block scales can be set with this scale_dict
            if not self.has_two_text_encoders and "text_encoder_2" in scale_dict:
                del scale_dict["text_encoder_2"]

            pipe.set_adapters("adapter-1", scale_dict)  # test will fail if this line throws an error

1183
    def test_simple_inference_with_text_denoiser_multi_adapter_delete_adapter(self):
1184
1185
1186
1187
        """
        Tests a simple inference with lora attached to text encoder and unet, attaches
        multiple adapters and set/delete them
        """
Aryan's avatar
Aryan committed
1188
        for scheduler_cls in self.scheduler_classes:
1189
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1190
1191
1192
1193
1194
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1195
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
1196

Aryan's avatar
Aryan committed
1197
1198
1199
1200
1201
1202
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
1203

Aryan's avatar
Aryan committed
1204
1205
1206
1207
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1208

1209
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1210
1211
1212
1213
1214
1215
1216
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
1217
1218

            pipe.set_adapters("adapter-1")
Aryan's avatar
Aryan committed
1219
            output_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1220
1221

            pipe.set_adapters("adapter-2")
Aryan's avatar
Aryan committed
1222
            output_adapter_2 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1223
1224

            pipe.set_adapters(["adapter-1", "adapter-2"])
Aryan's avatar
Aryan committed
1225
            output_adapter_mixed = pipe(**inputs, generator=torch.manual_seed(0))[0]
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242

            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
                "Adapter 1 and 2 should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 1 and mixed adapters should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_2, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 2 and mixed adapters should give different results",
            )

            pipe.delete_adapters("adapter-1")
Aryan's avatar
Aryan committed
1243
            output_deleted_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1244
1245
1246
1247
1248
1249
1250

            self.assertTrue(
                np.allclose(output_deleted_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
                "Adapter 1 and 2 should give different results",
            )

            pipe.delete_adapters("adapter-2")
Aryan's avatar
Aryan committed
1251
            output_deleted_adapters = pipe(**inputs, generator=torch.manual_seed(0))[0]
1252
1253
1254
1255
1256
1257

            self.assertTrue(
                np.allclose(output_no_lora, output_deleted_adapters, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

Aryan's avatar
Aryan committed
1258
1259
1260
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
1261

Aryan's avatar
Aryan committed
1262
1263
1264
1265
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1266
1267
1268
1269

            pipe.set_adapters(["adapter-1", "adapter-2"])
            pipe.delete_adapters(["adapter-1", "adapter-2"])

Aryan's avatar
Aryan committed
1270
            output_deleted_adapters = pipe(**inputs, generator=torch.manual_seed(0))[0]
1271
1272
1273
1274
1275
1276

            self.assertTrue(
                np.allclose(output_no_lora, output_deleted_adapters, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

1277
    def test_simple_inference_with_text_denoiser_multi_adapter_weighted(self):
1278
1279
1280
1281
        """
        Tests a simple inference with lora attached to text encoder and unet, attaches
        multiple adapters and set them
        """
Aryan's avatar
Aryan committed
1282
        for scheduler_cls in self.scheduler_classes:
1283
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1284
1285
1286
1287
1288
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1289
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
1290

Aryan's avatar
Aryan committed
1291
1292
1293
1294
1295
1296
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
1297

Aryan's avatar
Aryan committed
1298
1299
1300
1301
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1302

1303
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1304
1305
1306
1307
1308
1309
1310
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
1311
1312

            pipe.set_adapters("adapter-1")
Aryan's avatar
Aryan committed
1313
            output_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1314
1315

            pipe.set_adapters("adapter-2")
Aryan's avatar
Aryan committed
1316
            output_adapter_2 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1317
1318

            pipe.set_adapters(["adapter-1", "adapter-2"])
Aryan's avatar
Aryan committed
1319
            output_adapter_mixed = pipe(**inputs, generator=torch.manual_seed(0))[0]
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337

            # Fuse and unfuse should lead to the same results
            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
                "Adapter 1 and 2 should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_1, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 1 and mixed adapters should give different results",
            )

            self.assertFalse(
                np.allclose(output_adapter_2, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Adapter 2 and mixed adapters should give different results",
            )

            pipe.set_adapters(["adapter-1", "adapter-2"], [0.5, 0.6])
Aryan's avatar
Aryan committed
1338
            output_adapter_mixed_weighted = pipe(**inputs, generator=torch.manual_seed(0))[0]
1339
1340
1341
1342
1343
1344
1345

            self.assertFalse(
                np.allclose(output_adapter_mixed_weighted, output_adapter_mixed, atol=1e-3, rtol=1e-3),
                "Weighted adapter and mixed adapter should give different results",
            )

            pipe.disable_lora()
Aryan's avatar
Aryan committed
1346
            output_disabled = pipe(**inputs, generator=torch.manual_seed(0))[0]
1347
1348
1349
1350
1351
1352

            self.assertTrue(
                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
                "output with no lora and output with lora disabled should give same results",
            )

1353
    @skip_mps
1354
    def test_lora_fuse_nan(self):
Aryan's avatar
Aryan committed
1355
        for scheduler_cls in self.scheduler_classes:
1356
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1357
1358
1359
1360
1361
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1362
1363
1364
1365
1366
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )
1367

Aryan's avatar
Aryan committed
1368
1369
1370
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1371
1372
1373

            # corrupt one LoRA weight with `inf` values
            with torch.no_grad():
1374
1375
1376
1377
1378
1379
                if self.unet_kwargs:
                    pipe.unet.mid_block.attentions[0].transformer_blocks[0].attn1.to_q.lora_A[
                        "adapter-1"
                    ].weight += float("inf")
                else:
                    pipe.transformer.transformer_blocks[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
1380
1381
1382

            # with `safe_fusing=True` we should see an Error
            with self.assertRaises(ValueError):
Aryan's avatar
Aryan committed
1383
                pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=True)
1384
1385

            # without we should not see an error, but every image will be black
Aryan's avatar
Aryan committed
1386
1387
            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, safe_fusing=False)
            out = pipe("test", num_inference_steps=2, output_type="np")[0]
1388
1389
1390
1391
1392
1393
1394
1395

            self.assertTrue(np.isnan(out).all())

    def test_get_adapters(self):
        """
        Tests a simple usecase where we attach multiple adapters and check if the results
        are the expected results
        """
Aryan's avatar
Aryan committed
1396
        for scheduler_cls in self.scheduler_classes:
1397
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1398
1399
1400
1401
1402
1403
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

            pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
Aryan's avatar
Aryan committed
1404
1405
1406

            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
1407
1408
1409
1410
1411

            adapter_names = pipe.get_active_adapters()
            self.assertListEqual(adapter_names, ["adapter-1"])

            pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
Aryan's avatar
Aryan committed
1412
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424

            adapter_names = pipe.get_active_adapters()
            self.assertListEqual(adapter_names, ["adapter-2"])

            pipe.set_adapters(["adapter-1", "adapter-2"])
            self.assertListEqual(pipe.get_active_adapters(), ["adapter-1", "adapter-2"])

    def test_get_list_adapters(self):
        """
        Tests a simple usecase where we attach multiple adapters and check if the results
        are the expected results
        """
Aryan's avatar
Aryan committed
1425
        for scheduler_cls in self.scheduler_classes:
1426
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1427
1428
1429
1430
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)

Aryan's avatar
Aryan committed
1431
1432
1433
1434
1435
1436
            # 1.
            dicts_to_be_checked = {}
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                dicts_to_be_checked = {"text_encoder": ["adapter-1"]}

1437
1438
1439
1440
            if self.unet_kwargs is not None:
                pipe.unet.add_adapter(denoiser_lora_config, "adapter-1")
                dicts_to_be_checked.update({"unet": ["adapter-1"]})
            else:
Aryan's avatar
Aryan committed
1441
                pipe.transformer.add_adapter(denoiser_lora_config, "adapter-1")
1442
                dicts_to_be_checked.update({"transformer": ["adapter-1"]})
1443

Aryan's avatar
Aryan committed
1444
1445
1446
1447
1448
1449
1450
1451
            self.assertDictEqual(pipe.get_list_adapters(), dicts_to_be_checked)

            # 2.
            dicts_to_be_checked = {}
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
                dicts_to_be_checked = {"text_encoder": ["adapter-1", "adapter-2"]}

1452
1453
1454
1455
            if self.unet_kwargs is not None:
                pipe.unet.add_adapter(denoiser_lora_config, "adapter-2")
                dicts_to_be_checked.update({"unet": ["adapter-1", "adapter-2"]})
            else:
Aryan's avatar
Aryan committed
1456
                pipe.transformer.add_adapter(denoiser_lora_config, "adapter-2")
1457
                dicts_to_be_checked.update({"transformer": ["adapter-1", "adapter-2"]})
1458

Aryan's avatar
Aryan committed
1459
1460
1461
            self.assertDictEqual(pipe.get_list_adapters(), dicts_to_be_checked)

            # 3.
1462
            pipe.set_adapters(["adapter-1", "adapter-2"])
Aryan's avatar
Aryan committed
1463
1464
1465
1466
1467

            dicts_to_be_checked = {}
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                dicts_to_be_checked = {"text_encoder": ["adapter-1", "adapter-2"]}

1468
1469
1470
1471
            if self.unet_kwargs is not None:
                dicts_to_be_checked.update({"unet": ["adapter-1", "adapter-2"]})
            else:
                dicts_to_be_checked.update({"transformer": ["adapter-1", "adapter-2"]})
Aryan's avatar
Aryan committed
1472

1473
1474
            self.assertDictEqual(
                pipe.get_list_adapters(),
1475
                dicts_to_be_checked,
1476
1477
            )

Aryan's avatar
Aryan committed
1478
1479
1480
1481
1482
            # 4.
            dicts_to_be_checked = {}
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                dicts_to_be_checked = {"text_encoder": ["adapter-1", "adapter-2"]}

1483
            if self.unet_kwargs is not None:
Aryan's avatar
Aryan committed
1484
                pipe.unet.add_adapter(denoiser_lora_config, "adapter-3")
1485
1486
                dicts_to_be_checked.update({"unet": ["adapter-1", "adapter-2", "adapter-3"]})
            else:
Aryan's avatar
Aryan committed
1487
                pipe.transformer.add_adapter(denoiser_lora_config, "adapter-3")
1488
                dicts_to_be_checked.update({"transformer": ["adapter-1", "adapter-2", "adapter-3"]})
Aryan's avatar
Aryan committed
1489

1490
            self.assertDictEqual(pipe.get_list_adapters(), dicts_to_be_checked)
1491
1492

    @require_peft_version_greater(peft_version="0.6.2")
Aryan's avatar
Aryan committed
1493
1494
1495
    def test_simple_inference_with_text_lora_denoiser_fused_multi(
        self, expected_atol: float = 1e-3, expected_rtol: float = 1e-3
    ):
1496
1497
1498
1499
        """
        Tests a simple inference with lora attached into text encoder + fuses the lora weights into base model
        and makes sure it works as expected - with unet and multi-adapter case
        """
Aryan's avatar
Aryan committed
1500
        for scheduler_cls in self.scheduler_classes:
1501
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1502
1503
1504
1505
1506
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1507
            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
1508
            self.assertTrue(output_no_lora.shape == self.output_shape)
1509

Aryan's avatar
Aryan committed
1510
1511
1512
1513
1514
1515
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
                )

Aryan's avatar
Aryan committed
1516
1517
            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config, "adapter-1")
1518
1519

            # Attach a second adapter
Aryan's avatar
Aryan committed
1520
1521
1522
            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
                pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")

Aryan's avatar
Aryan committed
1523
            denoiser.add_adapter(denoiser_lora_config, "adapter-2")
1524

Aryan's avatar
Aryan committed
1525
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1526

1527
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1528
1529
1530
1531
1532
1533
1534
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
                    pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
1535
1536
1537

            # set them to multi-adapter inference mode
            pipe.set_adapters(["adapter-1", "adapter-2"])
Aryan's avatar
Aryan committed
1538
            outputs_all_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
1539
1540

            pipe.set_adapters(["adapter-1"])
Aryan's avatar
Aryan committed
1541
            outputs_lora_1 = pipe(**inputs, generator=torch.manual_seed(0))[0]
1542

Aryan's avatar
Aryan committed
1543
            pipe.fuse_lora(components=self.pipeline_class._lora_loadable_modules, adapter_names=["adapter-1"])
1544
1545

            # Fusing should still keep the LoRA layers so outpout should remain the same
Aryan's avatar
Aryan committed
1546
            outputs_lora_1_fused = pipe(**inputs, generator=torch.manual_seed(0))[0]
1547
1548

            self.assertTrue(
Aryan's avatar
Aryan committed
1549
                np.allclose(outputs_lora_1, outputs_lora_1_fused, atol=expected_atol, rtol=expected_rtol),
1550
1551
1552
                "Fused lora should not change the output",
            )

Aryan's avatar
Aryan committed
1553
1554
1555
1556
            pipe.unfuse_lora(components=self.pipeline_class._lora_loadable_modules)
            pipe.fuse_lora(
                components=self.pipeline_class._lora_loadable_modules, adapter_names=["adapter-2", "adapter-1"]
            )
1557
1558

            # Fusing should still keep the LoRA layers
Aryan's avatar
Aryan committed
1559
            output_all_lora_fused = pipe(**inputs, generator=torch.manual_seed(0))[0]
1560
            self.assertTrue(
Aryan's avatar
Aryan committed
1561
                np.allclose(output_all_lora_fused, outputs_all_lora, atol=expected_atol, rtol=expected_rtol),
1562
1563
1564
                "Fused lora should not change the output",
            )

1565
1566
    @require_peft_version_greater(peft_version="0.9.0")
    def test_simple_inference_with_dora(self):
Aryan's avatar
Aryan committed
1567
        for scheduler_cls in self.scheduler_classes:
1568
1569
1570
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(
                scheduler_cls, use_dora=True
            )
1571
1572
1573
1574
1575
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

Aryan's avatar
Aryan committed
1576
            output_no_dora_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
Sayak Paul's avatar
Sayak Paul committed
1577
            self.assertTrue(output_no_dora_lora.shape == self.output_shape)
1578
1579
1580

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
Aryan's avatar
Aryan committed
1581
1582
1583
1584

            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1585

1586
            if self.has_two_text_encoders or self.has_three_text_encoders:
Sayak Paul's avatar
Sayak Paul committed
1587
1588
1589
1590
1591
1592
                lora_loadable_components = self.pipeline_class._lora_loadable_modules
                if "text_encoder_2" in lora_loadable_components:
                    pipe.text_encoder_2.add_adapter(text_lora_config)
                    self.assertTrue(
                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                    )
1593

Aryan's avatar
Aryan committed
1594
            output_dora_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
1595
1596
1597
1598
1599
1600

            self.assertFalse(
                np.allclose(output_dora_lora, output_no_dora_lora, atol=1e-3, rtol=1e-3),
                "DoRA lora should change the output",
            )

1601
    @unittest.skip("This is failing for now - need to investigate")
1602
    def test_simple_inference_with_text_denoiser_lora_unfused_torch_compile(self):
1603
1604
1605
1606
        """
        Tests a simple inference with lora attached to text encoder and unet, then unloads the lora weights
        and makes sure it works as expected
        """
Aryan's avatar
Aryan committed
1607
        for scheduler_cls in self.scheduler_classes:
1608
            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
1609
1610
1611
1612
1613
1614
1615
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _, _, inputs = self.get_dummy_inputs(with_generator=False)

            pipe.text_encoder.add_adapter(text_lora_config)
            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
Aryan's avatar
Aryan committed
1616
1617
1618
1619

            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
            denoiser.add_adapter(denoiser_lora_config)
            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
1620

1621
            if self.has_two_text_encoders or self.has_three_text_encoders:
1622
1623
1624
1625
1626
1627
1628
1629
                pipe.text_encoder_2.add_adapter(text_lora_config)
                self.assertTrue(
                    check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
                )

            pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
            pipe.text_encoder = torch.compile(pipe.text_encoder, mode="reduce-overhead", fullgraph=True)

1630
            if self.has_two_text_encoders or self.has_three_text_encoders:
1631
1632
1633
                pipe.text_encoder_2 = torch.compile(pipe.text_encoder_2, mode="reduce-overhead", fullgraph=True)

            # Just makes sure it works..
Aryan's avatar
Aryan committed
1634
            _ = pipe(**inputs, generator=torch.manual_seed(0))[0]
1635
1636
1637
1638
1639
1640
1641

    def test_modify_padding_mode(self):
        def set_pad_mode(network, mode="circular"):
            for _, module in network.named_modules():
                if isinstance(module, torch.nn.Conv2d):
                    module.padding_mode = mode

Aryan's avatar
Aryan committed
1642
        for scheduler_cls in self.scheduler_classes:
1643
1644
1645
1646
1647
1648
1649
1650
1651
            components, _, _ = self.get_dummy_components(scheduler_cls)
            pipe = self.pipeline_class(**components)
            pipe = pipe.to(torch_device)
            pipe.set_progress_bar_config(disable=None)
            _pad_mode = "circular"
            set_pad_mode(pipe.vae, _pad_mode)
            set_pad_mode(pipe.unet, _pad_mode)

            _, _, inputs = self.get_dummy_inputs()
Aryan's avatar
Aryan committed
1652
            _ = pipe(**inputs)[0]