test_modeling_common.py 106 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2025 HuggingFace Inc.
3
4
5
6
7
8
9
10
11
12
13
14
15
#
# 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.

16
import copy
Aryan's avatar
Aryan committed
17
import gc
18
import glob
19
import inspect
20
21
import json
import os
Aryan's avatar
Aryan committed
22
import re
23
import tempfile
24
import traceback
25
import unittest
26
import unittest.mock as mock
27
import uuid
28
import warnings
29
30
from collections import defaultdict
from typing import Dict, List, Optional, Tuple, Union
31
32

import numpy as np
33
import pytest
34
import requests_mock
35
import safetensors.torch
36
import torch
37
import torch.nn as nn
YiYi Xu's avatar
YiYi Xu committed
38
from accelerate.utils.modeling import _get_proper_dtype, compute_module_sizes, dtype_byte_size
39
from huggingface_hub import ModelCard, delete_repo, snapshot_download, try_to_load_from_cache
40
from huggingface_hub.utils import HfHubHTTPError, is_jinja_available
41
from parameterized import parameterized
42

43
from diffusers.models import FluxTransformer2DModel, SD3Transformer2DModel, UNet2DConditionModel
44
45
46
47
48
49
from diffusers.models.attention_processor import (
    AttnProcessor,
    AttnProcessor2_0,
    AttnProcessorNPU,
    XFormersAttnProcessor,
)
hlky's avatar
hlky committed
50
from diffusers.models.auto_model import AutoModel
51
from diffusers.training_utils import EMAModel
52
53
54
from diffusers.utils import (
    SAFE_WEIGHTS_INDEX_NAME,
    WEIGHTS_INDEX_NAME,
55
    is_peft_available,
56
57
58
59
    is_torch_npu_available,
    is_xformers_available,
    logging,
)
60
from diffusers.utils.hub_utils import _add_variant
61
62
63
64
from diffusers.utils.torch_utils import get_torch_cuda_device_capability

from ..others.test_utils import TOKEN, USER, is_staging_test
from ..testing_utils import (
65
    CaptureLogger,
66
    _check_safetensors_serialization,
67
    backend_empty_cache,
68
69
70
    backend_max_memory_allocated,
    backend_reset_peak_memory_stats,
    backend_synchronize,
71
    check_if_dicts_are_equal,
72
    get_python_version,
73
    is_torch_compile,
Aryan's avatar
Aryan committed
74
    numpy_cosine_similarity_distance,
75
76
    require_peft_backend,
    require_peft_version_greater,
77
    require_torch_2,
78
    require_torch_accelerator,
Arsalan's avatar
Arsalan committed
79
    require_torch_accelerator_with_training,
80
    require_torch_multi_accelerator,
81
    require_torch_version_greater,
82
    run_test_in_subprocess,
83
    slow,
84
    torch_all_close,
Dhruv Nair's avatar
Dhruv Nair committed
85
    torch_device,
86
)
87
88


89
90
91
92
if is_peft_available():
    from peft.tuners.tuners_utils import BaseTunerLayer


93
94
95
96
97
98
99
100
101
def caculate_expected_num_shards(index_map_path):
    with open(index_map_path) as f:
        weight_map_dict = json.load(f)["weight_map"]
    first_key = list(weight_map_dict.keys())[0]
    weight_loc = weight_map_dict[first_key]  # e.g., diffusion_pytorch_model-00001-of-00002.safetensors
    expected_num_shards = int(weight_loc.split("-")[-1].split(".")[0])
    return expected_num_shards


102
103
104
105
106
107
108
109
110
111
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


112
113
114
115
116
117
118
119
120
121
122
# Will be run via run_test_in_subprocess
def _test_from_save_pretrained_dynamo(in_queue, out_queue, timeout):
    error = None
    try:
        init_dict, model_class = in_queue.get(timeout=timeout)

        model = model_class(**init_dict)
        model.to(torch_device)
        model = torch.compile(model)

        with tempfile.TemporaryDirectory() as tmpdirname:
123
            model.save_pretrained(tmpdirname, safe_serialization=False)
124
125
126
127
128
129
130
131
132
133
            new_model = model_class.from_pretrained(tmpdirname)
            new_model.to(torch_device)

        assert new_model.__class__ == model_class
    except Exception:
        error = f"{traceback.format_exc()}"

    results = {"error": error}
    out_queue.put(results, timeout=timeout)
    out_queue.join()
134
135


136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
def named_persistent_module_tensors(
    module: nn.Module,
    recurse: bool = False,
):
    """
    A helper function that gathers all the tensors (parameters + persistent buffers) of a given module.

    Args:
        module (`torch.nn.Module`):
            The module we want the tensors on.
        recurse (`bool`, *optional`, defaults to `False`):
            Whether or not to go look in every submodule or just return the direct parameters and buffers.
    """
    yield from module.named_parameters(recurse=recurse)

    for named_buffer in module.named_buffers(recurse=recurse):
        name, _ = named_buffer
        # Get parent by splitting on dots and traversing the model
        parent = module
        if "." in name:
            parent_name = name.rsplit(".", 1)[0]
            for part in parent_name.split("."):
                parent = getattr(parent, part)
            name = name.split(".")[-1]
        if name not in parent._non_persistent_buffers_set:
            yield named_buffer


def compute_module_persistent_sizes(
    model: nn.Module,
    dtype: Optional[Union[str, torch.device]] = None,
    special_dtypes: Optional[Dict[str, Union[str, torch.device]]] = None,
):
    """
    Compute the size of each submodule of a given model (parameters + persistent buffers).
    """
    if dtype is not None:
        dtype = _get_proper_dtype(dtype)
        dtype_size = dtype_byte_size(dtype)
    if special_dtypes is not None:
        special_dtypes = {key: _get_proper_dtype(dtyp) for key, dtyp in special_dtypes.items()}
        special_dtypes_size = {key: dtype_byte_size(dtyp) for key, dtyp in special_dtypes.items()}
    module_sizes = defaultdict(int)

    module_list = []

    module_list = named_persistent_module_tensors(model, recurse=True)

    for name, tensor in module_list:
        if special_dtypes is not None and name in special_dtypes:
            size = tensor.numel() * special_dtypes_size[name]
        elif dtype is None:
            size = tensor.numel() * dtype_byte_size(tensor.dtype)
        elif str(tensor.dtype).startswith(("torch.uint", "torch.int", "torch.bool")):
            # According to the code in set_module_tensor_to_device, these types won't be converted
            # so use their original size here
            size = tensor.numel() * dtype_byte_size(tensor.dtype)
        else:
            size = tensor.numel() * min(dtype_size, dtype_byte_size(tensor.dtype))
        name_parts = name.split(".")
        for idx in range(len(name_parts) + 1):
            module_sizes[".".join(name_parts[:idx])] += size

    return module_sizes


Aryan's avatar
Aryan committed
202
203
204
205
206
207
208
209
210
211
def cast_maybe_tensor_dtype(maybe_tensor, current_dtype, target_dtype):
    if torch.is_tensor(maybe_tensor):
        return maybe_tensor.to(target_dtype) if maybe_tensor.dtype == current_dtype else maybe_tensor
    if isinstance(maybe_tensor, dict):
        return {k: cast_maybe_tensor_dtype(v, current_dtype, target_dtype) for k, v in maybe_tensor.items()}
    if isinstance(maybe_tensor, list):
        return [cast_maybe_tensor_dtype(v, current_dtype, target_dtype) for v in maybe_tensor]
    return maybe_tensor


212
class ModelUtilsTest(unittest.TestCase):
213
214
215
    def tearDown(self):
        super().tearDown()

216
217
    def test_missing_key_loading_warning_message(self):
        with self.assertLogs("diffusers.models.modeling_utils", level="WARNING") as logs:
218
219
220
            UNet2DConditionModel.from_pretrained("hf-internal-testing/stable-diffusion-broken", subfolder="unet")

        # make sure that error message states what keys are missing
221
        assert "conv_out.bias" in " ".join(logs.output)
222

223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
    @parameterized.expand(
        [
            ("hf-internal-testing/tiny-stable-diffusion-pipe-variants-all-kinds", "unet", False),
            ("hf-internal-testing/tiny-stable-diffusion-pipe-variants-all-kinds", "unet", True),
            ("hf-internal-testing/tiny-sd-unet-with-sharded-ckpt", None, False),
            ("hf-internal-testing/tiny-sd-unet-with-sharded-ckpt", None, True),
        ]
    )
    def test_variant_sharded_ckpt_legacy_format_raises_warning(self, repo_id, subfolder, use_local):
        def load_model(path):
            kwargs = {"variant": "fp16"}
            if subfolder:
                kwargs["subfolder"] = subfolder
            return UNet2DConditionModel.from_pretrained(path, **kwargs)

        with self.assertWarns(FutureWarning) as warning:
            if use_local:
                with tempfile.TemporaryDirectory() as tmpdirname:
                    tmpdirname = snapshot_download(repo_id=repo_id)
                    _ = load_model(tmpdirname)
            else:
                _ = load_model(repo_id)

        warning_message = str(warning.warnings[0].message)
        self.assertIn("This serialization format is now deprecated to standardize the serialization", warning_message)

    # Local tests are already covered down below.
    @parameterized.expand(
        [
            ("hf-internal-testing/tiny-sd-unet-sharded-latest-format", None, "fp16"),
            ("hf-internal-testing/tiny-sd-unet-sharded-latest-format-subfolder", "unet", "fp16"),
            ("hf-internal-testing/tiny-sd-unet-sharded-no-variants", None, None),
            ("hf-internal-testing/tiny-sd-unet-sharded-no-variants-subfolder", "unet", None),
        ]
    )
    def test_variant_sharded_ckpt_loads_from_hub(self, repo_id, subfolder, variant=None):
        def load_model():
            kwargs = {}
            if variant:
                kwargs["variant"] = variant
            if subfolder:
                kwargs["subfolder"] = subfolder
            return UNet2DConditionModel.from_pretrained(repo_id, **kwargs)

        assert load_model()

269
270
271
272
273
    def test_cached_files_are_used_when_no_internet(self):
        # A mock response for an HTTP head request to emulate server down
        response_mock = mock.Mock()
        response_mock.status_code = 500
        response_mock.headers = {}
274
        response_mock.raise_for_status.side_effect = HfHubHTTPError("Server down", response=mock.Mock())
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
        response_mock.json.return_value = {}

        # Download this model to make sure it's in the cache.
        orig_model = UNet2DConditionModel.from_pretrained(
            "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet"
        )

        # Under the mock environment we get a 500 error when trying to reach the model.
        with mock.patch("requests.request", return_value=response_mock):
            # Download this model to make sure it's in the cache.
            model = UNet2DConditionModel.from_pretrained(
                "hf-internal-testing/tiny-stable-diffusion-torch", subfolder="unet", local_files_only=True
            )

        for p1, p2 in zip(orig_model.parameters(), model.parameters()):
            if p1.data.ne(p2.data).sum() > 0:
                assert False, "Parameters not the same!"

293
294
295
296
297
    def test_local_files_only_with_sharded_checkpoint(self):
        repo_id = "hf-internal-testing/tiny-flux-sharded"
        error_response = mock.Mock(
            status_code=500,
            headers={},
298
            raise_for_status=mock.Mock(side_effect=HfHubHTTPError("Server down", response=mock.Mock())),
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
            json=mock.Mock(return_value={}),
        )

        with tempfile.TemporaryDirectory() as tmpdir:
            model = FluxTransformer2DModel.from_pretrained(repo_id, subfolder="transformer", cache_dir=tmpdir)

            with mock.patch("requests.Session.get", return_value=error_response):
                # Should fail with local_files_only=False (network required)
                # We would make a network call with model_info
                with self.assertRaises(OSError):
                    FluxTransformer2DModel.from_pretrained(
                        repo_id, subfolder="transformer", cache_dir=tmpdir, local_files_only=False
                    )

                # Should succeed with local_files_only=True (uses cache)
                # model_info call skipped
                local_model = FluxTransformer2DModel.from_pretrained(
                    repo_id, subfolder="transformer", cache_dir=tmpdir, local_files_only=True
                )

            assert all(torch.equal(p1, p2) for p1, p2 in zip(model.parameters(), local_model.parameters())), (
                "Model parameters don't match!"
            )

            # Remove a shard file
            cached_shard_file = try_to_load_from_cache(
                repo_id, filename="transformer/diffusion_pytorch_model-00001-of-00002.safetensors", cache_dir=tmpdir
            )
            os.remove(cached_shard_file)

            # Attempting to load from cache should raise an error
            with self.assertRaises(OSError) as context:
                FluxTransformer2DModel.from_pretrained(
                    repo_id, subfolder="transformer", cache_dir=tmpdir, local_files_only=True
                )

            # Verify error mentions the missing shard
            error_msg = str(context.exception)
            assert cached_shard_file in error_msg or "required according to the checkpoint index" in error_msg, (
                f"Expected error about missing shard, got: {error_msg}"
            )

341
    @unittest.skip("Flaky behaviour on CI. Re-enable after migrating to new runners")
342
    @unittest.skipIf(torch_device == "mps", reason="Test not supported for MPS.")
343
    def test_one_request_upon_cached(self):
344
        use_safetensors = False
345
346
347
348

        with tempfile.TemporaryDirectory() as tmpdirname:
            with requests_mock.mock(real_http=True) as m:
                UNet2DConditionModel.from_pretrained(
349
350
351
352
                    "hf-internal-testing/tiny-stable-diffusion-torch",
                    subfolder="unet",
                    cache_dir=tmpdirname,
                    use_safetensors=use_safetensors,
353
354
355
                )

            download_requests = [r.method for r in m.request_history]
356
357
358
            assert download_requests.count("HEAD") == 3, (
                "3 HEAD requests one for config, one for model, and one for shard index file."
            )
359
360
361
362
            assert download_requests.count("GET") == 2, "2 GET requests one for config, one for model"

            with requests_mock.mock(real_http=True) as m:
                UNet2DConditionModel.from_pretrained(
363
364
365
366
                    "hf-internal-testing/tiny-stable-diffusion-torch",
                    subfolder="unet",
                    cache_dir=tmpdirname,
                    use_safetensors=use_safetensors,
367
368
369
                )

            cache_requests = [r.method for r in m.request_history]
370
371
372
            assert "HEAD" == cache_requests[0] and len(cache_requests) == 2, (
                "We should call only `model_info` to check for commit hash and  knowing if shard index is present."
            )
373

374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
    def test_weight_overwrite(self):
        with tempfile.TemporaryDirectory() as tmpdirname, self.assertRaises(ValueError) as error_context:
            UNet2DConditionModel.from_pretrained(
                "hf-internal-testing/tiny-stable-diffusion-torch",
                subfolder="unet",
                cache_dir=tmpdirname,
                in_channels=9,
            )

        # make sure that error message states what keys are missing
        assert "Cannot load" in str(error_context.exception)

        with tempfile.TemporaryDirectory() as tmpdirname:
            model = UNet2DConditionModel.from_pretrained(
                "hf-internal-testing/tiny-stable-diffusion-torch",
                subfolder="unet",
                cache_dir=tmpdirname,
                in_channels=9,
                low_cpu_mem_usage=False,
                ignore_mismatched_sizes=True,
            )

        assert model.config.in_channels == 9

398
    @require_torch_accelerator
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
    def test_keep_modules_in_fp32(self):
        r"""
        A simple tests to check if the modules under `_keep_in_fp32_modules` are kept in fp32 when we load the model in fp16/bf16
        Also ensures if inference works.
        """
        fp32_modules = SD3Transformer2DModel._keep_in_fp32_modules

        for torch_dtype in [torch.bfloat16, torch.float16]:
            SD3Transformer2DModel._keep_in_fp32_modules = ["proj_out"]

            model = SD3Transformer2DModel.from_pretrained(
                "hf-internal-testing/tiny-sd3-pipe", subfolder="transformer", torch_dtype=torch_dtype
            ).to(torch_device)

            for name, module in model.named_modules():
                if isinstance(module, torch.nn.Linear):
                    if name in model._keep_in_fp32_modules:
                        self.assertTrue(module.weight.dtype == torch.float32)
                    else:
                        self.assertTrue(module.weight.dtype == torch_dtype)

        def get_dummy_inputs():
            batch_size = 2
            num_channels = 4
            height = width = embedding_dim = 32
            pooled_embedding_dim = embedding_dim * 2
            sequence_length = 154

            hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device)
            encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
            pooled_prompt_embeds = torch.randn((batch_size, pooled_embedding_dim)).to(torch_device)
            timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)

            return {
                "hidden_states": hidden_states,
                "encoder_hidden_states": encoder_hidden_states,
                "pooled_projections": pooled_prompt_embeds,
                "timestep": timestep,
            }

        # test if inference works.
        with torch.no_grad() and torch.amp.autocast(torch_device, dtype=torch_dtype):
            input_dict_for_transformer = get_dummy_inputs()
            model_inputs = {
                k: v.to(device=torch_device) for k, v in input_dict_for_transformer.items() if not isinstance(v, bool)
            }
            model_inputs.update({k: v for k, v in input_dict_for_transformer.items() if k not in model_inputs})
            _ = model(**model_inputs)

        SD3Transformer2DModel._keep_in_fp32_modules = fp32_modules

450

451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
class UNetTesterMixin:
    def test_forward_with_norm_groups(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        init_dict["norm_num_groups"] = 16
        init_dict["block_out_channels"] = (16, 32)

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

            if isinstance(output, dict):
                output = output.to_tuple()[0]

        self.assertIsNotNone(output)
        expected_shape = inputs_dict["sample"].shape
        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")


473
class ModelTesterMixin:
474
475
    main_input_name = None  # overwrite in model specific tester class
    base_precision = 1e-3
Will Berman's avatar
Will Berman committed
476
    forward_requires_fresh_args = False
477
    model_split_percents = [0.5, 0.7, 0.9]
478
    uses_custom_attn_processor = False
479
480
481
482
483
484
485
486
487
488
489
490
491
492

    def check_device_map_is_respected(self, model, device_map):
        for param_name, param in model.named_parameters():
            # Find device in device_map
            while len(param_name) > 0 and param_name not in device_map:
                param_name = ".".join(param_name.split(".")[:-1])
            if param_name not in device_map:
                raise ValueError("device map is incomplete, it does not contain any device for `param_name`.")

            param_device = device_map[param_name]
            if param_device in ["cpu", "disk"]:
                self.assertEqual(param.device, torch.device("meta"))
            else:
                self.assertEqual(param.device, torch.device(param_device))
493

494
    def test_from_save_pretrained(self, expected_max_diff=5e-5):
Will Berman's avatar
Will Berman committed
495
496
497
498
499
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
500

501
502
        if hasattr(model, "set_default_attn_processor"):
            model.set_default_attn_processor()
503
504
505
506
        model.to(torch_device)
        model.eval()

        with tempfile.TemporaryDirectory() as tmpdirname:
507
            model.save_pretrained(tmpdirname, safe_serialization=False)
508
            new_model = self.model_class.from_pretrained(tmpdirname)
509
510
            if hasattr(new_model, "set_default_attn_processor"):
                new_model.set_default_attn_processor()
511
512
513
            new_model.to(torch_device)

        with torch.no_grad():
Will Berman's avatar
Will Berman committed
514
515
516
517
518
            if self.forward_requires_fresh_args:
                image = model(**self.inputs_dict(0))
            else:
                image = model(**inputs_dict)

519
            if isinstance(image, dict):
520
                image = image.to_tuple()[0]
521

Will Berman's avatar
Will Berman committed
522
523
524
525
            if self.forward_requires_fresh_args:
                new_image = new_model(**self.inputs_dict(0))
            else:
                new_image = new_model(**inputs_dict)
526
527

            if isinstance(new_image, dict):
528
                new_image = new_image.to_tuple()[0]
529

530
531
        max_diff = (image - new_image).abs().max().item()
        self.assertLessEqual(max_diff, expected_max_diff, "Models give different forward passes")
532

533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
    def test_getattr_is_correct(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)

        # save some things to test
        model.dummy_attribute = 5
        model.register_to_config(test_attribute=5)

        logger = logging.get_logger("diffusers.models.modeling_utils")
        # 30 for warning
        logger.setLevel(30)
        with CaptureLogger(logger) as cap_logger:
            assert hasattr(model, "dummy_attribute")
            assert getattr(model, "dummy_attribute") == 5
            assert model.dummy_attribute == 5

        # no warning should be thrown
        assert cap_logger.out == ""

        logger = logging.get_logger("diffusers.models.modeling_utils")
        # 30 for warning
        logger.setLevel(30)
        with CaptureLogger(logger) as cap_logger:
            assert hasattr(model, "save_pretrained")
            fn = model.save_pretrained
            fn_1 = getattr(model, "save_pretrained")

            assert fn == fn_1
        # no warning should be thrown
        assert cap_logger.out == ""

        # warning should be thrown
        with self.assertWarns(FutureWarning):
            assert model.test_attribute == 5

        with self.assertWarns(FutureWarning):
            assert getattr(model, "test_attribute") == 5

        with self.assertRaises(AttributeError) as error:
            model.does_not_exist

        assert str(error.exception) == f"'{type(model).__name__}' object has no attribute 'does_not_exist'"

576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
    @unittest.skipIf(
        torch_device != "npu" or not is_torch_npu_available(),
        reason="torch npu flash attention is only available with NPU and `torch_npu` installed",
    )
    def test_set_torch_npu_flash_attn_processor_determinism(self):
        torch.use_deterministic_algorithms(False)
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
        model.to(torch_device)

        if not hasattr(model, "set_attn_processor"):
            # If not has `set_attn_processor`, skip test
            return

        model.set_default_attn_processor()
        assert all(type(proc) == AttnProcessorNPU for proc in model.attn_processors.values())
        with torch.no_grad():
            if self.forward_requires_fresh_args:
                output = model(**self.inputs_dict(0))[0]
            else:
                output = model(**inputs_dict)[0]

        model.enable_npu_flash_attention()
        assert all(type(proc) == AttnProcessorNPU for proc in model.attn_processors.values())
        with torch.no_grad():
            if self.forward_requires_fresh_args:
                output_2 = model(**self.inputs_dict(0))[0]
            else:
                output_2 = model(**inputs_dict)[0]

        model.set_attn_processor(AttnProcessorNPU())
        assert all(type(proc) == AttnProcessorNPU for proc in model.attn_processors.values())
        with torch.no_grad():
            if self.forward_requires_fresh_args:
                output_3 = model(**self.inputs_dict(0))[0]
            else:
                output_3 = model(**inputs_dict)[0]

        torch.use_deterministic_algorithms(True)

        assert torch.allclose(output, output_2, atol=self.base_precision)
        assert torch.allclose(output, output_3, atol=self.base_precision)
        assert torch.allclose(output_2, output_3, atol=self.base_precision)

Dhruv Nair's avatar
Dhruv Nair committed
623
624
625
626
627
628
    @unittest.skipIf(
        torch_device != "cuda" or not is_xformers_available(),
        reason="XFormers attention is only available with CUDA and `xformers` installed",
    )
    def test_set_xformers_attn_processor_for_determinism(self):
        torch.use_deterministic_algorithms(False)
Will Berman's avatar
Will Berman committed
629
630
631
632
633
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
Dhruv Nair's avatar
Dhruv Nair committed
634
635
636
637
        model.to(torch_device)

        if not hasattr(model, "set_attn_processor"):
            # If not has `set_attn_processor`, skip test
Dhruv Nair's avatar
Dhruv Nair committed
638
639
640
641
            return

        if not hasattr(model, "set_default_attn_processor"):
            # If not has `set_attn_processor`, skip test
Dhruv Nair's avatar
Dhruv Nair committed
642
643
644
645
646
            return

        model.set_default_attn_processor()
        assert all(type(proc) == AttnProcessor for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
647
648
649
650
            if self.forward_requires_fresh_args:
                output = model(**self.inputs_dict(0))[0]
            else:
                output = model(**inputs_dict)[0]
Dhruv Nair's avatar
Dhruv Nair committed
651
652
653
654

        model.enable_xformers_memory_efficient_attention()
        assert all(type(proc) == XFormersAttnProcessor for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
655
656
657
658
            if self.forward_requires_fresh_args:
                output_2 = model(**self.inputs_dict(0))[0]
            else:
                output_2 = model(**inputs_dict)[0]
Dhruv Nair's avatar
Dhruv Nair committed
659

660
661
662
        model.set_attn_processor(XFormersAttnProcessor())
        assert all(type(proc) == XFormersAttnProcessor for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
663
664
665
666
            if self.forward_requires_fresh_args:
                output_3 = model(**self.inputs_dict(0))[0]
            else:
                output_3 = model(**inputs_dict)[0]
667
668
669

        torch.use_deterministic_algorithms(True)

Dhruv Nair's avatar
Dhruv Nair committed
670
        assert torch.allclose(output, output_2, atol=self.base_precision)
671
672
        assert torch.allclose(output, output_3, atol=self.base_precision)
        assert torch.allclose(output_2, output_3, atol=self.base_precision)
Dhruv Nair's avatar
Dhruv Nair committed
673

674
    @require_torch_accelerator
675
    def test_set_attn_processor_for_determinism(self):
676
677
678
        if self.uses_custom_attn_processor:
            return

679
        torch.use_deterministic_algorithms(False)
Will Berman's avatar
Will Berman committed
680
681
682
683
684
685
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)

686
687
688
689
690
691
692
693
        model.to(torch_device)

        if not hasattr(model, "set_attn_processor"):
            # If not has `set_attn_processor`, skip test
            return

        assert all(type(proc) == AttnProcessor2_0 for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
694
695
696
697
            if self.forward_requires_fresh_args:
                output_1 = model(**self.inputs_dict(0))[0]
            else:
                output_1 = model(**inputs_dict)[0]
698
699
700
701

        model.set_default_attn_processor()
        assert all(type(proc) == AttnProcessor for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
702
703
704
705
            if self.forward_requires_fresh_args:
                output_2 = model(**self.inputs_dict(0))[0]
            else:
                output_2 = model(**inputs_dict)[0]
706
707
708
709

        model.set_attn_processor(AttnProcessor2_0())
        assert all(type(proc) == AttnProcessor2_0 for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
710
711
712
713
            if self.forward_requires_fresh_args:
                output_4 = model(**self.inputs_dict(0))[0]
            else:
                output_4 = model(**inputs_dict)[0]
714
715
716
717

        model.set_attn_processor(AttnProcessor())
        assert all(type(proc) == AttnProcessor for proc in model.attn_processors.values())
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
718
719
720
721
            if self.forward_requires_fresh_args:
                output_5 = model(**self.inputs_dict(0))[0]
            else:
                output_5 = model(**inputs_dict)[0]
722
723
724
725
726
727
728
729

        torch.use_deterministic_algorithms(True)

        # make sure that outputs match
        assert torch.allclose(output_2, output_1, atol=self.base_precision)
        assert torch.allclose(output_2, output_4, atol=self.base_precision)
        assert torch.allclose(output_2, output_5, atol=self.base_precision)

730
    def test_from_save_pretrained_variant(self, expected_max_diff=5e-5):
Will Berman's avatar
Will Berman committed
731
732
733
734
735
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
736

737
738
        if hasattr(model, "set_default_attn_processor"):
            model.set_default_attn_processor()
739

740
741
742
743
        model.to(torch_device)
        model.eval()

        with tempfile.TemporaryDirectory() as tmpdirname:
744
            model.save_pretrained(tmpdirname, variant="fp16", safe_serialization=False)
745
            new_model = self.model_class.from_pretrained(tmpdirname, variant="fp16")
746
747
            if hasattr(new_model, "set_default_attn_processor"):
                new_model.set_default_attn_processor()
748
749
750
751
752
753
754
755
756
757
758

            # non-variant cannot be loaded
            with self.assertRaises(OSError) as error_context:
                self.model_class.from_pretrained(tmpdirname)

            # make sure that error message states what keys are missing
            assert "Error no file named diffusion_pytorch_model.bin found in directory" in str(error_context.exception)

            new_model.to(torch_device)

        with torch.no_grad():
Will Berman's avatar
Will Berman committed
759
760
761
762
            if self.forward_requires_fresh_args:
                image = model(**self.inputs_dict(0))
            else:
                image = model(**inputs_dict)
763
            if isinstance(image, dict):
764
                image = image.to_tuple()[0]
765

Will Berman's avatar
Will Berman committed
766
767
768
769
            if self.forward_requires_fresh_args:
                new_image = new_model(**self.inputs_dict(0))
            else:
                new_image = new_model(**inputs_dict)
770
771

            if isinstance(new_image, dict):
772
                new_image = new_image.to_tuple()[0]
773

774
775
        max_diff = (image - new_image).abs().max().item()
        self.assertLessEqual(max_diff, expected_max_diff, "Models give different forward passes")
776

777
    @is_torch_compile
778
    @require_torch_2
779
780
781
782
    @unittest.skipIf(
        get_python_version == (3, 12),
        reason="Torch Dynamo isn't yet supported for Python 3.12.",
    )
783
    def test_from_save_pretrained_dynamo(self):
784
785
786
        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        inputs = [init_dict, self.model_class]
        run_test_in_subprocess(test_case=self, target_func=_test_from_save_pretrained_dynamo, inputs=inputs)
787

788
789
790
791
792
793
794
795
796
797
798
799
    def test_from_save_pretrained_dtype(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        for dtype in [torch.float32, torch.float16, torch.bfloat16]:
            if torch_device == "mps" and dtype == torch.bfloat16:
                continue
            with tempfile.TemporaryDirectory() as tmpdirname:
                model.to(dtype)
800
                model.save_pretrained(tmpdirname, safe_serialization=False)
801
                new_model = self.model_class.from_pretrained(tmpdirname, low_cpu_mem_usage=True, torch_dtype=dtype)
802
                assert new_model.dtype == dtype
803
804
805
806
807
808
809
810
                if (
                    hasattr(self.model_class, "_keep_in_fp32_modules")
                    and self.model_class._keep_in_fp32_modules is None
                ):
                    new_model = self.model_class.from_pretrained(
                        tmpdirname, low_cpu_mem_usage=False, torch_dtype=dtype
                    )
                    assert new_model.dtype == dtype
811

812
    def test_determinism(self, expected_max_diff=1e-5):
Will Berman's avatar
Will Berman committed
813
814
815
816
817
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
818
819
        model.to(torch_device)
        model.eval()
820

821
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
822
823
824
825
            if self.forward_requires_fresh_args:
                first = model(**self.inputs_dict(0))
            else:
                first = model(**inputs_dict)
826
            if isinstance(first, dict):
827
                first = first.to_tuple()[0]
828

Will Berman's avatar
Will Berman committed
829
830
831
832
            if self.forward_requires_fresh_args:
                second = model(**self.inputs_dict(0))
            else:
                second = model(**inputs_dict)
833
            if isinstance(second, dict):
834
                second = second.to_tuple()[0]
835
836
837
838
839
840

        out_1 = first.cpu().numpy()
        out_2 = second.cpu().numpy()
        out_1 = out_1[~np.isnan(out_1)]
        out_2 = out_2[~np.isnan(out_2)]
        max_diff = np.amax(np.abs(out_1 - out_2))
841
        self.assertLessEqual(max_diff, expected_max_diff)
842

843
    def test_output(self, expected_output_shape=None):
844
845
846
847
848
849
850
851
852
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        with torch.no_grad():
            output = model(**inputs_dict)

            if isinstance(output, dict):
853
                output = output.to_tuple()[0]
854
855

        self.assertIsNotNone(output)
856

857
858
        # input & output have to have the same shape
        input_tensor = inputs_dict[self.main_input_name]
859
860
861
862
863
864

        if expected_output_shape is None:
            expected_shape = input_tensor.shape
            self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
        else:
            self.assertEqual(output.shape, expected_output_shape, "Input and output shapes do not match")
865

866
    def test_model_from_pretrained(self):
867
868
869
870
871
872
873
874
875
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.eval()

        # test if the model can be loaded from the config
        # and has all the expected shape
        with tempfile.TemporaryDirectory() as tmpdirname:
876
            model.save_pretrained(tmpdirname, safe_serialization=False)
877
            new_model = self.model_class.from_pretrained(tmpdirname)
878
879
880
            new_model.to(torch_device)
            new_model.eval()

881
        # check if all parameters shape are the same
882
883
884
885
886
887
888
889
890
        for param_name in model.state_dict().keys():
            param_1 = model.state_dict()[param_name]
            param_2 = new_model.state_dict()[param_name]
            self.assertEqual(param_1.shape, param_2.shape)

        with torch.no_grad():
            output_1 = model(**inputs_dict)

            if isinstance(output_1, dict):
891
                output_1 = output_1.to_tuple()[0]
892
893
894
895

            output_2 = new_model(**inputs_dict)

            if isinstance(output_2, dict):
896
                output_2 = output_2.to_tuple()[0]
897
898
899

        self.assertEqual(output_1.shape, output_2.shape)

Arsalan's avatar
Arsalan committed
900
    @require_torch_accelerator_with_training
901
902
903
904
905
906
907
908
909
    def test_training(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.train()
        output = model(**inputs_dict)

        if isinstance(output, dict):
910
            output = output.to_tuple()[0]
911

912
913
        input_tensor = inputs_dict[self.main_input_name]
        noise = torch.randn((input_tensor.shape[0],) + self.output_shape).to(torch_device)
914
915
916
        loss = torch.nn.functional.mse_loss(output, noise)
        loss.backward()

Arsalan's avatar
Arsalan committed
917
    @require_torch_accelerator_with_training
918
919
920
921
922
923
    def test_ema_training(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict)
        model.to(torch_device)
        model.train()
924
        ema_model = EMAModel(model.parameters())
925
926
927
928

        output = model(**inputs_dict)

        if isinstance(output, dict):
929
            output = output.to_tuple()[0]
930

931
932
        input_tensor = inputs_dict[self.main_input_name]
        noise = torch.randn((input_tensor.shape[0],) + self.output_shape).to(torch_device)
933
934
        loss = torch.nn.functional.mse_loss(output, noise)
        loss.backward()
935
        ema_model.step(model.parameters())
936

937
    def test_outputs_equivalence(self):
938
        def set_nan_tensor_to_zero(t):
939
940
941
942
943
            # Temporary fallback until `aten::_index_put_impl_` is implemented in mps
            # Track progress in https://github.com/pytorch/pytorch/issues/77764
            device = t.device
            if device.type == "mps":
                t = t.to("cpu")
944
            t[t != t] = 0
945
            return t.to(device)
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968

        def recursive_check(tuple_object, dict_object):
            if isinstance(tuple_object, (List, Tuple)):
                for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object.values()):
                    recursive_check(tuple_iterable_value, dict_iterable_value)
            elif isinstance(tuple_object, Dict):
                for tuple_iterable_value, dict_iterable_value in zip(tuple_object.values(), dict_object.values()):
                    recursive_check(tuple_iterable_value, dict_iterable_value)
            elif tuple_object is None:
                return
            else:
                self.assertTrue(
                    torch.allclose(
                        set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                    ),
                    msg=(
                        "Tuple and dict output are not equal. Difference:"
                        f" {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`:"
                        f" {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has"
                        f" `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}."
                    ),
                )

Will Berman's avatar
Will Berman committed
969
970
971
972
973
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)
974
975
976
977

        model.to(torch_device)
        model.eval()

978
        with torch.no_grad():
Will Berman's avatar
Will Berman committed
979
980
981
982
983
984
            if self.forward_requires_fresh_args:
                outputs_dict = model(**self.inputs_dict(0))
                outputs_tuple = model(**self.inputs_dict(0), return_dict=False)
            else:
                outputs_dict = model(**inputs_dict)
                outputs_tuple = model(**inputs_dict, return_dict=False)
985
986

        recursive_check(outputs_tuple, outputs_dict)
987

Arsalan's avatar
Arsalan committed
988
    @require_torch_accelerator_with_training
989
    def test_enable_disable_gradient_checkpointing(self):
990
        # Skip test if model does not support gradient checkpointing
991
        if not self.model_class._supports_gradient_checkpointing:
992
            pytest.skip("Gradient checkpointing is not supported.")
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()

        # at init model should have gradient checkpointing disabled
        model = self.model_class(**init_dict)
        self.assertFalse(model.is_gradient_checkpointing)

        # check enable works
        model.enable_gradient_checkpointing()
        self.assertTrue(model.is_gradient_checkpointing)

        # check disable works
        model.disable_gradient_checkpointing()
        self.assertFalse(model.is_gradient_checkpointing)
1007

1008
    @require_torch_accelerator_with_training
1009
    def test_effective_gradient_checkpointing(self, loss_tolerance=1e-5, param_grad_tol=5e-5, skip: set[str] = {}):
1010
        # Skip test if model does not support gradient checkpointing
1011
        if not self.model_class._supports_gradient_checkpointing:
1012
            pytest.skip("Gradient checkpointing is not supported.")
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056

        # enable deterministic behavior for gradient checkpointing
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        inputs_dict_copy = copy.deepcopy(inputs_dict)
        torch.manual_seed(0)
        model = self.model_class(**init_dict)
        model.to(torch_device)

        assert not model.is_gradient_checkpointing and model.training

        out = model(**inputs_dict).sample
        # run the backwards pass on the model. For backwards pass, for simplicity purpose,
        # we won't calculate the loss and rather backprop on out.sum()
        model.zero_grad()

        labels = torch.randn_like(out)
        loss = (out - labels).mean()
        loss.backward()

        # re-instantiate the model now enabling gradient checkpointing
        torch.manual_seed(0)
        model_2 = self.model_class(**init_dict)
        # clone model
        model_2.load_state_dict(model.state_dict())
        model_2.to(torch_device)
        model_2.enable_gradient_checkpointing()

        assert model_2.is_gradient_checkpointing and model_2.training

        out_2 = model_2(**inputs_dict_copy).sample
        # run the backwards pass on the model. For backwards pass, for simplicity purpose,
        # we won't calculate the loss and rather backprop on out.sum()
        model_2.zero_grad()
        loss_2 = (out_2 - labels).mean()
        loss_2.backward()

        # compare the output and parameters gradients
        self.assertTrue((loss - loss_2).abs() < loss_tolerance)
        named_params = dict(model.named_parameters())
        named_params_2 = dict(model_2.named_parameters())

        for name, param in named_params.items():
            if "post_quant_conv" in name:
                continue
1057
1058
            if name in skip:
                continue
1059
1060
1061
1062
            # TODO(aryan): remove the below lines after looking into easyanimate transformer a little more
            # It currently errors out the gradient checkpointing test because the gradients for attn2.to_out is None
            if param.grad is None:
                continue
1063
1064
1065
1066
1067
1068
            self.assertTrue(torch_all_close(param.grad.data, named_params_2[name].grad.data, atol=param_grad_tol))

    @unittest.skipIf(torch_device == "mps", "This test is not supported for MPS devices.")
    def test_gradient_checkpointing_is_applied(
        self, expected_set=None, attention_head_dim=None, num_attention_heads=None, block_out_channels=None
    ):
1069
        # Skip test if model does not support gradient checkpointing
1070
        if not self.model_class._supports_gradient_checkpointing:
1071
            pytest.skip("Gradient checkpointing is not supported.")
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        if attention_head_dim is not None:
            init_dict["attention_head_dim"] = attention_head_dim
        if num_attention_heads is not None:
            init_dict["num_attention_heads"] = num_attention_heads
        if block_out_channels is not None:
            init_dict["block_out_channels"] = block_out_channels

        model_class_copy = copy.copy(self.model_class)
        model = model_class_copy(**init_dict)
        model.enable_gradient_checkpointing()

1086
1087
1088
1089
1090
1091
        modules_with_gc_enabled = {}
        for submodule in model.modules():
            if hasattr(submodule, "gradient_checkpointing"):
                self.assertTrue(submodule.gradient_checkpointing)
                modules_with_gc_enabled[submodule.__class__.__name__] = True

1092
1093
1094
        assert set(modules_with_gc_enabled.keys()) == expected_set
        assert all(modules_with_gc_enabled.values()), "All modules should be enabled"

1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
    def test_deprecated_kwargs(self):
        has_kwarg_in_model_class = "kwargs" in inspect.signature(self.model_class.__init__).parameters
        has_deprecated_kwarg = len(self.model_class._deprecated_kwargs) > 0

        if has_kwarg_in_model_class and not has_deprecated_kwarg:
            raise ValueError(
                f"{self.model_class} has `**kwargs` in its __init__ method but has not defined any deprecated kwargs"
                " under the `_deprecated_kwargs` class attribute. Make sure to either remove `**kwargs` if there are"
                " no deprecated arguments or add the deprecated argument with `_deprecated_kwargs ="
                " [<deprecated_argument>]`"
            )

        if not has_kwarg_in_model_class and has_deprecated_kwarg:
            raise ValueError(
                f"{self.model_class} doesn't have `**kwargs` in its __init__ method but has defined deprecated kwargs"
                " under the `_deprecated_kwargs` class attribute. Make sure to either add the `**kwargs` argument to"
                f" {self.model_class}.__init__ if there are deprecated arguments or remove the deprecated argument"
                " from `_deprecated_kwargs = [<deprecated_argument>]`"
            )
1114

1115
    @parameterized.expand([(4, 4, True), (4, 8, False), (8, 4, False)])
1116
1117
    @torch.no_grad()
    @unittest.skipIf(not is_peft_available(), "Only with PEFT")
1118
    def test_save_load_lora_adapter(self, rank, lora_alpha, use_dora=False):
1119
1120
1121
1122
1123
1124
1125
1126
1127
        from peft import LoraConfig
        from peft.utils import get_peft_model_state_dict

        from diffusers.loaders.peft import PeftAdapterMixin

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)

        if not issubclass(model.__class__, PeftAdapterMixin):
1128
            pytest.skip(f"PEFT is not supported for this model ({model.__class__.__name__}).")
1129
1130
1131
1132
1133

        torch.manual_seed(0)
        output_no_lora = model(**inputs_dict, return_dict=False)[0]

        denoiser_lora_config = LoraConfig(
1134
1135
            r=rank,
            lora_alpha=lora_alpha,
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
            init_lora_weights=False,
            use_dora=use_dora,
        )
        model.add_adapter(denoiser_lora_config)
        self.assertTrue(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

        torch.manual_seed(0)
        outputs_with_lora = model(**inputs_dict, return_dict=False)[0]

        self.assertFalse(torch.allclose(output_no_lora, outputs_with_lora, atol=1e-4, rtol=1e-4))

        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_lora_adapter(tmpdir)
            self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.safetensors")))

            state_dict_loaded = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))

            model.unload_lora()
            self.assertFalse(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

            model.load_lora_adapter(tmpdir, prefix=None, use_safetensors=True)
            state_dict_retrieved = get_peft_model_state_dict(model, adapter_name="default_0")

            for k in state_dict_loaded:
                loaded_v = state_dict_loaded[k]
                retrieved_v = state_dict_retrieved[k].to(loaded_v.device)
                self.assertTrue(torch.allclose(loaded_v, retrieved_v))

            self.assertTrue(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

        torch.manual_seed(0)
        outputs_with_lora_2 = model(**inputs_dict, return_dict=False)[0]

        self.assertFalse(torch.allclose(output_no_lora, outputs_with_lora_2, atol=1e-4, rtol=1e-4))
        self.assertTrue(torch.allclose(outputs_with_lora, outputs_with_lora_2, atol=1e-4, rtol=1e-4))

    @unittest.skipIf(not is_peft_available(), "Only with PEFT")
1174
    def test_lora_wrong_adapter_name_raises_error(self):
1175
1176
1177
1178
1179
1180
1181
1182
        from peft import LoraConfig

        from diffusers.loaders.peft import PeftAdapterMixin

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)

        if not issubclass(model.__class__, PeftAdapterMixin):
1183
            pytest.skip(f"PEFT is not supported for this model ({model.__class__.__name__}).")
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201

        denoiser_lora_config = LoraConfig(
            r=4,
            lora_alpha=4,
            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
            init_lora_weights=False,
            use_dora=False,
        )
        model.add_adapter(denoiser_lora_config)
        self.assertTrue(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

        with tempfile.TemporaryDirectory() as tmpdir:
            wrong_name = "foo"
            with self.assertRaises(ValueError) as err_context:
                model.save_lora_adapter(tmpdir, adapter_name=wrong_name)

            self.assertTrue(f"Adapter name {wrong_name} not found in the model." in str(err_context.exception))

1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
    @parameterized.expand([(4, 4, True), (4, 8, False), (8, 4, False)])
    @torch.no_grad()
    @unittest.skipIf(not is_peft_available(), "Only with PEFT")
    def test_lora_adapter_metadata_is_loaded_correctly(self, rank, lora_alpha, use_dora):
        from peft import LoraConfig

        from diffusers.loaders.peft import PeftAdapterMixin

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)

        if not issubclass(model.__class__, PeftAdapterMixin):
1214
            pytest.skip(f"PEFT is not supported for this model ({model.__class__.__name__}).")
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250

        denoiser_lora_config = LoraConfig(
            r=rank,
            lora_alpha=lora_alpha,
            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
            init_lora_weights=False,
            use_dora=use_dora,
        )
        model.add_adapter(denoiser_lora_config)
        metadata = model.peft_config["default"].to_dict()
        self.assertTrue(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_lora_adapter(tmpdir)
            model_file = os.path.join(tmpdir, "pytorch_lora_weights.safetensors")
            self.assertTrue(os.path.isfile(model_file))

            model.unload_lora()
            self.assertFalse(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

            model.load_lora_adapter(tmpdir, prefix=None, use_safetensors=True)
            parsed_metadata = model.peft_config["default_0"].to_dict()
            check_if_dicts_are_equal(metadata, parsed_metadata)

    @torch.no_grad()
    @unittest.skipIf(not is_peft_available(), "Only with PEFT")
    def test_lora_adapter_wrong_metadata_raises_error(self):
        from peft import LoraConfig

        from diffusers.loaders.lora_base import LORA_ADAPTER_METADATA_KEY
        from diffusers.loaders.peft import PeftAdapterMixin

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)

        if not issubclass(model.__class__, PeftAdapterMixin):
1251
            pytest.skip(f"PEFT is not supported for this model ({model.__class__.__name__}).")
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285

        denoiser_lora_config = LoraConfig(
            r=4,
            lora_alpha=4,
            target_modules=["to_q", "to_k", "to_v", "to_out.0"],
            init_lora_weights=False,
            use_dora=False,
        )
        model.add_adapter(denoiser_lora_config)
        self.assertTrue(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_lora_adapter(tmpdir)
            model_file = os.path.join(tmpdir, "pytorch_lora_weights.safetensors")
            self.assertTrue(os.path.isfile(model_file))

            # Perturb the metadata in the state dict.
            loaded_state_dict = safetensors.torch.load_file(model_file)
            metadata = {"format": "pt"}
            lora_adapter_metadata = denoiser_lora_config.to_dict()
            lora_adapter_metadata.update({"foo": 1, "bar": 2})
            for key, value in lora_adapter_metadata.items():
                if isinstance(value, set):
                    lora_adapter_metadata[key] = list(value)
            metadata[LORA_ADAPTER_METADATA_KEY] = json.dumps(lora_adapter_metadata, indent=2, sort_keys=True)
            safetensors.torch.save_file(loaded_state_dict, model_file, metadata=metadata)

            model.unload_lora()
            self.assertFalse(check_if_lora_correctly_set(model), "LoRA layers not set correctly")

            with self.assertRaises(TypeError) as err_context:
                model.load_lora_adapter(tmpdir, prefix=None, use_safetensors=True)
            self.assertTrue("`LoraConfig` class could not be instantiated" in str(err_context.exception))

1286
    @require_torch_accelerator
1287
    def test_cpu_offload(self):
1288
1289
        if self.model_class._no_split_modules is None:
            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
1290
1291
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1292

1293
1294
1295
1296
1297
        model = model.to(torch_device)

        torch.manual_seed(0)
        base_output = model(**inputs_dict)

YiYi Xu's avatar
YiYi Xu committed
1298
        model_size = compute_module_sizes(model)[""]
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
        # We test several splits of sizes to make sure it works.
        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir)

            for max_size in max_gpu_sizes:
                max_memory = {0: max_size, "cpu": model_size * 2}
                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                # Making sure part of the model will actually end up offloaded
                self.assertSetEqual(set(new_model.hf_device_map.values()), {0, "cpu"})

                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
                torch.manual_seed(0)
                new_output = new_model(**inputs_dict)

                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1316
    @require_torch_accelerator
1317
    def test_disk_offload_without_safetensors(self):
1318
1319
        if self.model_class._no_split_modules is None:
            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
1320
1321
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1322

1323
1324
1325
1326
1327
        model = model.to(torch_device)

        torch.manual_seed(0)
        base_output = model(**inputs_dict)

YiYi Xu's avatar
YiYi Xu committed
1328
        model_size = compute_module_sizes(model)[""]
1329
1330
1331
1332
        max_size = int(self.model_split_percents[0] * model_size)
        # Force disk offload by setting very small CPU memory
        max_memory = {0: max_size, "cpu": int(0.1 * max_size)}

1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir, safe_serialization=False)
            with self.assertRaises(ValueError):
                # This errors out because it's missing an offload folder
                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)

            new_model = self.model_class.from_pretrained(
                tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
            )

            self.check_device_map_is_respected(new_model, new_model.hf_device_map)
            torch.manual_seed(0)
            new_output = new_model(**inputs_dict)

            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1349
    @require_torch_accelerator
1350
    def test_disk_offload_with_safetensors(self):
1351
1352
        if self.model_class._no_split_modules is None:
            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
1353
1354
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1355

1356
1357
1358
1359
1360
        model = model.to(torch_device)

        torch.manual_seed(0)
        base_output = model(**inputs_dict)

YiYi Xu's avatar
YiYi Xu committed
1361
        model_size = compute_module_sizes(model)[""]
1362
1363
1364
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir)

1365
            max_size = int(self.model_split_percents[0] * model_size)
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
            max_memory = {0: max_size, "cpu": max_size}
            new_model = self.model_class.from_pretrained(
                tmp_dir, device_map="auto", offload_folder=tmp_dir, max_memory=max_memory
            )

            self.check_device_map_is_respected(new_model, new_model.hf_device_map)
            torch.manual_seed(0)
            new_output = new_model(**inputs_dict)

            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1377
    @require_torch_multi_accelerator
1378
    def test_model_parallelism(self):
1379
1380
        if self.model_class._no_split_modules is None:
            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
1381
1382
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1383

1384
1385
1386
1387
1388
        model = model.to(torch_device)

        torch.manual_seed(0)
        base_output = model(**inputs_dict)

YiYi Xu's avatar
YiYi Xu committed
1389
        model_size = compute_module_sizes(model)[""]
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
        # We test several splits of sizes to make sure it works.
        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir)

            for max_size in max_gpu_sizes:
                max_memory = {0: max_size, 1: model_size * 2, "cpu": model_size * 2}
                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
                # Making sure part of the model will actually end up offloaded
                self.assertSetEqual(set(new_model.hf_device_map.values()), {0, 1})

                self.check_device_map_is_respected(new_model, new_model.hf_device_map)

                torch.manual_seed(0)
                new_output = new_model(**inputs_dict)

                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1408
    @require_torch_accelerator
1409
    def test_sharded_checkpoints(self):
1410
        torch.manual_seed(0)
1411
1412
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1413
1414
1415
1416
        model = model.to(torch_device)

        base_output = model(**inputs_dict)

1417
        model_size = compute_module_persistent_sizes(model)[""]
1418
1419
1420
1421
1422
1423
1424
1425
        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small.
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir, max_shard_size=f"{max_shard_size}KB")
            self.assertTrue(os.path.exists(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))

            # Now check if the right number of shards exists. First, let's get the number of shards.
            # Since this number can be dependent on the model being tested, it's important that we calculate it
            # instead of hardcoding it.
1426
            expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME))
1427
1428
1429
            actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(".safetensors")])
            self.assertTrue(actual_num_shards == expected_num_shards)

1430
            new_model = self.model_class.from_pretrained(tmp_dir).eval()
1431
            new_model = new_model.to(torch_device)
1432
1433

            torch.manual_seed(0)
1434
1435
            if "generator" in inputs_dict:
                _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
1436
            new_output = new_model(**inputs_dict)
1437

1438
1439
            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1440
    @require_torch_accelerator
1441
1442
1443
1444
1445
1446
1447
1448
    def test_sharded_checkpoints_with_variant(self):
        torch.manual_seed(0)
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
        model = model.to(torch_device)

        base_output = model(**inputs_dict)

1449
        model_size = compute_module_persistent_sizes(model)[""]
1450
1451
1452
1453
1454
1455
1456
        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small.
        variant = "fp16"
        with tempfile.TemporaryDirectory() as tmp_dir:
            # It doesn't matter if the actual model is in fp16 or not. Just adding the variant and
            # testing if loading works with the variant when the checkpoint is sharded should be
            # enough.
            model.cpu().save_pretrained(tmp_dir, max_shard_size=f"{max_shard_size}KB", variant=variant)
1457

1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
            index_filename = _add_variant(SAFE_WEIGHTS_INDEX_NAME, variant)
            self.assertTrue(os.path.exists(os.path.join(tmp_dir, index_filename)))

            # Now check if the right number of shards exists. First, let's get the number of shards.
            # Since this number can be dependent on the model being tested, it's important that we calculate it
            # instead of hardcoding it.
            expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, index_filename))
            actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(".safetensors")])
            self.assertTrue(actual_num_shards == expected_num_shards)

            new_model = self.model_class.from_pretrained(tmp_dir, variant=variant).eval()
            new_model = new_model.to(torch_device)

            torch.manual_seed(0)
            if "generator" in inputs_dict:
                _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            new_output = new_model(**inputs_dict)

            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
    @require_torch_accelerator
    def test_sharded_checkpoints_with_parallel_loading(self):
        torch.manual_seed(0)
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
        model = model.to(torch_device)

        base_output = model(**inputs_dict)

        model_size = compute_module_persistent_sizes(model)[""]
        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small.
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir, max_shard_size=f"{max_shard_size}KB")
            self.assertTrue(os.path.exists(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))

            # Now check if the right number of shards exists. First, let's get the number of shards.
            # Since this number can be dependent on the model being tested, it's important that we calculate it
            # instead of hardcoding it.
            expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME))
            actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(".safetensors")])
            self.assertTrue(actual_num_shards == expected_num_shards)

            # Load with parallel loading
            os.environ["HF_ENABLE_PARALLEL_LOADING"] = "yes"
            new_model = self.model_class.from_pretrained(tmp_dir).eval()
            new_model = new_model.to(torch_device)

            torch.manual_seed(0)
            if "generator" in inputs_dict:
                _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            new_output = new_model(**inputs_dict)
            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
            # set to no.
            os.environ["HF_ENABLE_PARALLEL_LOADING"] = "no"

1513
    @require_torch_accelerator
1514
    def test_sharded_checkpoints_device_map(self):
1515
1516
        if self.model_class._no_split_modules is None:
            pytest.skip("Test not supported for this model as `_no_split_modules` is not set.")
1517
1518
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**config).eval()
1519
1520
1521
1522
1523
        model = model.to(torch_device)

        torch.manual_seed(0)
        base_output = model(**inputs_dict)

1524
        model_size = compute_module_persistent_sizes(model)[""]
1525
1526
1527
1528
1529
1530
1531
1532
        max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small.
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.cpu().save_pretrained(tmp_dir, max_shard_size=f"{max_shard_size}KB")
            self.assertTrue(os.path.exists(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))

            # Now check if the right number of shards exists. First, let's get the number of shards.
            # Since this number can be dependent on the model being tested, it's important that we calculate it
            # instead of hardcoding it.
1533
            expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME))
1534
1535
1536
1537
1538
1539
            actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(".safetensors")])
            self.assertTrue(actual_num_shards == expected_num_shards)

            new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto")

            torch.manual_seed(0)
1540
1541
            if "generator" in inputs_dict:
                _, inputs_dict = self.prepare_init_args_and_inputs_for_common()
1542
1543
1544
            new_output = new_model(**inputs_dict)
            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))

1545
1546
1547
1548
1549
1550
1551
1552
    # This test is okay without a GPU because we're not running any execution. We're just serializing
    # and check if the resultant files are following an expected format.
    def test_variant_sharded_ckpt_right_format(self):
        for use_safe in [True, False]:
            extension = ".safetensors" if use_safe else ".bin"
            config, _ = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**config).eval()

1553
            model_size = compute_module_persistent_sizes(model)[""]
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
            max_shard_size = int((model_size * 0.75) / (2**10))  # Convert to KB as these test models are small.
            variant = "fp16"
            with tempfile.TemporaryDirectory() as tmp_dir:
                model.cpu().save_pretrained(
                    tmp_dir, variant=variant, max_shard_size=f"{max_shard_size}KB", safe_serialization=use_safe
                )
                index_variant = _add_variant(SAFE_WEIGHTS_INDEX_NAME if use_safe else WEIGHTS_INDEX_NAME, variant)
                self.assertTrue(os.path.exists(os.path.join(tmp_dir, index_variant)))

                # Now check if the right number of shards exists. First, let's get the number of shards.
                # Since this number can be dependent on the model being tested, it's important that we calculate it
                # instead of hardcoding it.
                expected_num_shards = caculate_expected_num_shards(os.path.join(tmp_dir, index_variant))
                actual_num_shards = len([file for file in os.listdir(tmp_dir) if file.endswith(extension)])
                self.assertTrue(actual_num_shards == expected_num_shards)

                # Check if the variant is present as a substring in the checkpoints.
                shard_files = [
                    file
                    for file in os.listdir(tmp_dir)
                    if file.endswith(extension) or ("index" in file and "json" in file)
                ]
                assert all(variant in f for f in shard_files)

                # Check if the sharded checkpoints were serialized in the right format.
                shard_files = [file for file in os.listdir(tmp_dir) if file.endswith(extension)]
                # Example: diffusion_pytorch_model.fp16-00001-of-00002.safetensors
                assert all(f.split(".")[1].split("-")[0] == variant for f in shard_files)

1583
1584
1585
    def test_layerwise_casting_training(self):
        def test_fn(storage_dtype, compute_dtype):
            if torch.device(torch_device).type == "cpu" and compute_dtype == torch.bfloat16:
1586
                pytest.skip("Skipping test because CPU doesn't go well with bfloat16.")
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

            model = self.model_class(**init_dict)
            model = model.to(torch_device, dtype=compute_dtype)
            model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)
            model.train()

            inputs_dict = cast_maybe_tensor_dtype(inputs_dict, torch.float32, compute_dtype)
            with torch.amp.autocast(device_type=torch.device(torch_device).type):
                output = model(**inputs_dict)

                if isinstance(output, dict):
                    output = output.to_tuple()[0]

                input_tensor = inputs_dict[self.main_input_name]
                noise = torch.randn((input_tensor.shape[0],) + self.output_shape).to(torch_device)
                noise = cast_maybe_tensor_dtype(noise, torch.float32, compute_dtype)
                loss = torch.nn.functional.mse_loss(output, noise)

            loss.backward()

        test_fn(torch.float16, torch.float32)
        test_fn(torch.float8_e4m3fn, torch.float32)
        test_fn(torch.float8_e5m2, torch.float32)
        test_fn(torch.float8_e4m3fn, torch.bfloat16)

1613
    @torch.no_grad()
Aryan's avatar
Aryan committed
1614
    def test_layerwise_casting_inference(self):
1615
1616
        from diffusers.hooks._common import _GO_LC_SUPPORTED_PYTORCH_LAYERS
        from diffusers.hooks.layerwise_casting import DEFAULT_SKIP_MODULES_PATTERN
Aryan's avatar
Aryan committed
1617
1618
1619

        torch.manual_seed(0)
        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
1620
1621
1622
1623
        model = self.model_class(**config)
        model.eval()
        model.to(torch_device)
        base_slice = model(**inputs_dict)[0].detach().flatten().cpu().numpy()
Aryan's avatar
Aryan committed
1624
1625
1626
1627
1628
1629

        def check_linear_dtype(module, storage_dtype, compute_dtype):
            patterns_to_check = DEFAULT_SKIP_MODULES_PATTERN
            if getattr(module, "_skip_layerwise_casting_patterns", None) is not None:
                patterns_to_check += tuple(module._skip_layerwise_casting_patterns)
            for name, submodule in module.named_modules():
1630
                if not isinstance(submodule, _GO_LC_SUPPORTED_PYTORCH_LAYERS):
Aryan's avatar
Aryan committed
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
                    continue
                dtype_to_check = storage_dtype
                if any(re.search(pattern, name) for pattern in patterns_to_check):
                    dtype_to_check = compute_dtype
                if getattr(submodule, "weight", None) is not None:
                    self.assertEqual(submodule.weight.dtype, dtype_to_check)
                if getattr(submodule, "bias", None) is not None:
                    self.assertEqual(submodule.bias.dtype, dtype_to_check)

        def test_layerwise_casting(storage_dtype, compute_dtype):
            torch.manual_seed(0)
            config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            inputs_dict = cast_maybe_tensor_dtype(inputs_dict, torch.float32, compute_dtype)
            model = self.model_class(**config).eval()
            model = model.to(torch_device, dtype=compute_dtype)
            model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)

            check_linear_dtype(model, storage_dtype, compute_dtype)
            output = model(**inputs_dict)[0].float().flatten().detach().cpu().numpy()

            # The precision test is not very important for fast tests. In most cases, the outputs will not be the same.
            # We just want to make sure that the layerwise casting is working as expected.
            self.assertTrue(numpy_cosine_similarity_distance(base_slice, output) < 1.0)

        test_layerwise_casting(torch.float16, torch.float32)
        test_layerwise_casting(torch.float8_e4m3fn, torch.float32)
        test_layerwise_casting(torch.float8_e5m2, torch.float32)
        test_layerwise_casting(torch.float8_e4m3fn, torch.bfloat16)

1660
    @require_torch_accelerator
1661
    @torch.no_grad()
Aryan's avatar
Aryan committed
1662
1663
    def test_layerwise_casting_memory(self):
        MB_TOLERANCE = 0.2
1664
        LEAST_COMPUTE_CAPABILITY = 8.0
Aryan's avatar
Aryan committed
1665
1666
1667

        def reset_memory_stats():
            gc.collect()
1668
1669
1670
            backend_synchronize(torch_device)
            backend_empty_cache(torch_device)
            backend_reset_peak_memory_stats(torch_device)
Aryan's avatar
Aryan committed
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682

        def get_memory_usage(storage_dtype, compute_dtype):
            torch.manual_seed(0)
            config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            inputs_dict = cast_maybe_tensor_dtype(inputs_dict, torch.float32, compute_dtype)
            model = self.model_class(**config).eval()
            model = model.to(torch_device, dtype=compute_dtype)
            model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)

            reset_memory_stats()
            model(**inputs_dict)
            model_memory_footprint = model.get_memory_footprint()
1683
            peak_inference_memory_allocated_mb = backend_max_memory_allocated(torch_device) / 1024**2
Aryan's avatar
Aryan committed
1684
1685
1686
1687
1688
1689
1690
1691
1692

            return model_memory_footprint, peak_inference_memory_allocated_mb

        fp32_memory_footprint, fp32_max_memory = get_memory_usage(torch.float32, torch.float32)
        fp8_e4m3_fp32_memory_footprint, fp8_e4m3_fp32_max_memory = get_memory_usage(torch.float8_e4m3fn, torch.float32)
        fp8_e4m3_bf16_memory_footprint, fp8_e4m3_bf16_max_memory = get_memory_usage(
            torch.float8_e4m3fn, torch.bfloat16
        )

1693
        compute_capability = get_torch_cuda_device_capability() if torch_device == "cuda" else None
Aryan's avatar
Aryan committed
1694
        self.assertTrue(fp8_e4m3_bf16_memory_footprint < fp8_e4m3_fp32_memory_footprint < fp32_memory_footprint)
1695
1696
1697
1698
        # NOTE: the following assertion would fail on our CI (running Tesla T4) due to bf16 using more memory than fp32.
        # On other devices, such as DGX (Ampere) and Audace (Ada), the test passes. So, we conditionally check it.
        if compute_capability and compute_capability >= LEAST_COMPUTE_CAPABILITY:
            self.assertTrue(fp8_e4m3_bf16_max_memory < fp8_e4m3_fp32_max_memory)
Aryan's avatar
Aryan committed
1699
1700
1701
1702
1703
1704
1705
1706
        # On this dummy test case with a small model, sometimes fp8_e4m3_fp32 max memory usage is higher than fp32 by a few
        # bytes. This only happens for some models, so we allow a small tolerance.
        # For any real model being tested, the order would be fp8_e4m3_bf16 < fp8_e4m3_fp32 < fp32.
        self.assertTrue(
            fp8_e4m3_fp32_max_memory < fp32_max_memory
            or abs(fp8_e4m3_fp32_max_memory - fp32_max_memory) < MB_TOLERANCE
        )

1707
    @parameterized.expand([False, True])
1708
    @require_torch_accelerator
1709
    def test_group_offloading(self, record_stream):
1710
1711
1712
        if not self.model_class._supports_group_offloading:
            pytest.skip("Model does not support group offloading.")

Aryan's avatar
Aryan committed
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        torch.manual_seed(0)

        @torch.no_grad()
        def run_forward(model):
            self.assertTrue(
                all(
                    module._diffusers_hook.get_hook("group_offloading") is not None
                    for module in model.modules()
                    if hasattr(module, "_diffusers_hook")
                )
            )
            model.eval()
            return model(**inputs_dict)[0]

        model = self.model_class(**init_dict)

        model.to(torch_device)
        output_without_group_offloading = run_forward(model)

        torch.manual_seed(0)
        model = self.model_class(**init_dict)
        model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1)
        output_with_group_offloading1 = run_forward(model)

        torch.manual_seed(0)
        model = self.model_class(**init_dict)
        model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, non_blocking=True)
        output_with_group_offloading2 = run_forward(model)

        torch.manual_seed(0)
        model = self.model_class(**init_dict)
        model.enable_group_offload(torch_device, offload_type="leaf_level")
        output_with_group_offloading3 = run_forward(model)

        torch.manual_seed(0)
        model = self.model_class(**init_dict)
1750
1751
1752
        model.enable_group_offload(
            torch_device, offload_type="leaf_level", use_stream=True, record_stream=record_stream
        )
Aryan's avatar
Aryan committed
1753
1754
1755
1756
1757
1758
1759
        output_with_group_offloading4 = run_forward(model)

        self.assertTrue(torch.allclose(output_without_group_offloading, output_with_group_offloading1, atol=1e-5))
        self.assertTrue(torch.allclose(output_without_group_offloading, output_with_group_offloading2, atol=1e-5))
        self.assertTrue(torch.allclose(output_without_group_offloading, output_with_group_offloading3, atol=1e-5))
        self.assertTrue(torch.allclose(output_without_group_offloading, output_with_group_offloading4, atol=1e-5))

1760
1761
1762
1763
    @parameterized.expand([(False, "block_level"), (True, "leaf_level")])
    @require_torch_accelerator
    @torch.no_grad()
    def test_group_offloading_with_layerwise_casting(self, record_stream, offload_type):
1764
1765
1766
        if not self.model_class._supports_group_offloading:
            pytest.skip("Model does not support group offloading.")

1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
        torch.manual_seed(0)
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)

        model.to(torch_device)
        model.eval()
        _ = model(**inputs_dict)[0]

        torch.manual_seed(0)
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        storage_dtype, compute_dtype = torch.float16, torch.float32
        inputs_dict = cast_maybe_tensor_dtype(inputs_dict, torch.float32, compute_dtype)
        model = self.model_class(**init_dict)
        model.eval()
        additional_kwargs = {} if offload_type == "leaf_level" else {"num_blocks_per_group": 1}
        model.enable_group_offload(
            torch_device, offload_type=offload_type, use_stream=True, record_stream=record_stream, **additional_kwargs
        )
        model.enable_layerwise_casting(storage_dtype=storage_dtype, compute_dtype=compute_dtype)
        _ = model(**inputs_dict)[0]

1788
    @parameterized.expand([("block_level", False), ("leaf_level", True)])
1789
1790
    @require_torch_accelerator
    @torch.no_grad()
1791
1792
    @torch.inference_mode()
    def test_group_offloading_with_disk(self, offload_type, record_stream, atol=1e-5):
1793
1794
1795
        if not self.model_class._supports_group_offloading:
            pytest.skip("Model does not support group offloading.")

1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
        def _has_generator_arg(model):
            sig = inspect.signature(model.forward)
            params = sig.parameters
            return "generator" in params

        def _run_forward(model, inputs_dict):
            accepts_generator = _has_generator_arg(model)
            if accepts_generator:
                inputs_dict["generator"] = torch.manual_seed(0)
            torch.manual_seed(0)
            return model(**inputs_dict)[0]

        if self.__class__.__name__ == "AutoencoderKLCosmosTests" and offload_type == "leaf_level":
            pytest.skip("With `leaf_type` as the offloading type, it fails. Needs investigation.")

1811
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
1812
        torch.manual_seed(0)
1813
        model = self.model_class(**init_dict)
1814

1815
        model.eval()
1816
1817
1818
1819
1820
1821
1822
1823
1824
        model.to(torch_device)
        output_without_group_offloading = _run_forward(model, inputs_dict)

        torch.manual_seed(0)
        model = self.model_class(**init_dict)
        model.eval()

        num_blocks_per_group = None if offload_type == "leaf_level" else 1
        additional_kwargs = {} if offload_type == "leaf_level" else {"num_blocks_per_group": num_blocks_per_group}
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
        with tempfile.TemporaryDirectory() as tmpdir:
            model.enable_group_offload(
                torch_device,
                offload_type=offload_type,
                offload_to_disk_path=tmpdir,
                use_stream=True,
                record_stream=record_stream,
                **additional_kwargs,
            )
            has_safetensors = glob.glob(f"{tmpdir}/*.safetensors")
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
            self.assertTrue(has_safetensors, "No safetensors found in the directory.")

            # For "leaf-level", there is a prefetching hook which makes this check a bit non-deterministic
            # in nature. So, skip it.
            if offload_type != "leaf_level":
                is_correct, extra_files, missing_files = _check_safetensors_serialization(
                    module=model,
                    offload_to_disk_path=tmpdir,
                    offload_type=offload_type,
                    num_blocks_per_group=num_blocks_per_group,
                )
                if not is_correct:
                    if extra_files:
                        raise ValueError(f"Found extra files: {', '.join(extra_files)}")
                    elif missing_files:
                        raise ValueError(f"Following files are missing: {', '.join(missing_files)}")

            output_with_group_offloading = _run_forward(model, inputs_dict)
            self.assertTrue(torch.allclose(output_without_group_offloading, output_with_group_offloading, atol=atol))
1854

hlky's avatar
hlky committed
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
    def test_auto_model(self, expected_max_diff=5e-5):
        if self.forward_requires_fresh_args:
            model = self.model_class(**self.init_dict)
        else:
            init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict)

        model = model.eval()
        model = model.to(torch_device)

        if hasattr(model, "set_default_attn_processor"):
            model.set_default_attn_processor()

        with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmpdirname:
            model.save_pretrained(tmpdirname, safe_serialization=False)

            auto_model = AutoModel.from_pretrained(tmpdirname)
            if hasattr(auto_model, "set_default_attn_processor"):
                auto_model.set_default_attn_processor()

        auto_model = auto_model.eval()
        auto_model = auto_model.to(torch_device)

        with torch.no_grad():
            if self.forward_requires_fresh_args:
                output_original = model(**self.inputs_dict(0))
                output_auto = auto_model(**self.inputs_dict(0))
            else:
                output_original = model(**inputs_dict)
                output_auto = auto_model(**inputs_dict)

            if isinstance(output_original, dict):
                output_original = output_original.to_tuple()[0]
            if isinstance(output_auto, dict):
                output_auto = output_auto.to_tuple()[0]

        max_diff = (output_original - output_auto).abs().max().item()
        self.assertLessEqual(
            max_diff,
            expected_max_diff,
            f"AutoModel forward pass diff: {max_diff} exceeds threshold {expected_max_diff}",
        )

1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
    @parameterized.expand(
        [
            (-1, "You can't pass device_map as a negative int"),
            ("foo", "When passing device_map as a string, the value needs to be a device name"),
        ]
    )
    def test_wrong_device_map_raises_error(self, device_map, msg_substring):
        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_pretrained(tmpdir)
            with self.assertRaises(ValueError) as err_ctx:
                _ = self.model_class.from_pretrained(tmpdir, device_map=device_map)

        assert msg_substring in str(err_ctx.exception)

1914
1915
    @parameterized.expand([0, torch_device, torch.device(torch_device)])
    @require_torch_accelerator
1916
1917
1918
1919
1920
1921
1922
1923
    def test_passing_non_dict_device_map_works(self, device_map):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).eval()
        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_pretrained(tmpdir)
            loaded_model = self.model_class.from_pretrained(tmpdir, device_map=device_map)
            _ = loaded_model(**inputs_dict)

1924
1925
    @parameterized.expand([("", torch_device), ("", torch.device(torch_device))])
    @require_torch_accelerator
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
    def test_passing_dict_device_map_works(self, name, device):
        # There are other valid dict-based `device_map` values too. It's best to refer to
        # the docs for those: https://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference#the-devicemap.
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).eval()
        device_map = {name: device}
        with tempfile.TemporaryDirectory() as tmpdir:
            model.save_pretrained(tmpdir)
            loaded_model = self.model_class.from_pretrained(tmpdir, device_map=device_map)
            _ = loaded_model(**inputs_dict)

1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004

@is_staging_test
class ModelPushToHubTester(unittest.TestCase):
    identifier = uuid.uuid4()
    repo_id = f"test-model-{identifier}"
    org_repo_id = f"valid_org/{repo_id}-org"

    def test_push_to_hub(self):
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        model.push_to_hub(self.repo_id, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(f"{USER}/{self.repo_id}")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=TOKEN, repo_id=self.repo_id)

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, repo_id=self.repo_id, push_to_hub=True, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(f"{USER}/{self.repo_id}")
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(self.repo_id, token=TOKEN)

    def test_push_to_hub_in_organization(self):
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        model.push_to_hub(self.org_repo_id, token=TOKEN)

        new_model = UNet2DConditionModel.from_pretrained(self.org_repo_id)
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(token=TOKEN, repo_id=self.org_repo_id)

        # Push to hub via save_pretrained
        with tempfile.TemporaryDirectory() as tmp_dir:
            model.save_pretrained(tmp_dir, push_to_hub=True, token=TOKEN, repo_id=self.org_repo_id)

        new_model = UNet2DConditionModel.from_pretrained(self.org_repo_id)
        for p1, p2 in zip(model.parameters(), new_model.parameters()):
            self.assertTrue(torch.equal(p1, p2))

        # Reset repo
        delete_repo(self.org_repo_id, token=TOKEN)
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027

    @unittest.skipIf(
        not is_jinja_available(),
        reason="Model card tests cannot be performed without Jinja installed.",
    )
    def test_push_to_hub_library_name(self):
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        model.push_to_hub(self.repo_id, token=TOKEN)

        model_card = ModelCard.load(f"{USER}/{self.repo_id}", token=TOKEN).data
        assert model_card.library_name == "diffusers"

        # Reset repo
        delete_repo(self.repo_id, token=TOKEN)
2028
2029


2030
@require_torch_accelerator
2031
2032
2033
@require_torch_2
@is_torch_compile
@slow
2034
@require_torch_version_greater("2.7.1")
2035
class TorchCompileTesterMixin:
2036
2037
    different_shapes_for_compilation = None

2038
2039
2040
    def setUp(self):
        # clean up the VRAM before each test
        super().setUp()
2041
        torch.compiler.reset()
2042
2043
2044
2045
2046
2047
        gc.collect()
        backend_empty_cache(torch_device)

    def tearDown(self):
        # clean up the VRAM after each test in case of CUDA runtime errors
        super().tearDown()
2048
        torch.compiler.reset()
2049
2050
2051
2052
2053
2054
2055
        gc.collect()
        backend_empty_cache(torch_device)

    def test_torch_compile_recompilation_and_graph_break(self):
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict).to(torch_device)
2056
        model.eval()
2057
2058
        model = torch.compile(model, fullgraph=True)

2059
2060
2061
2062
2063
        with (
            torch._inductor.utils.fresh_inductor_cache(),
            torch._dynamo.config.patch(error_on_recompile=True),
            torch.no_grad(),
        ):
2064
2065
            _ = model(**inputs_dict)
            _ = model(**inputs_dict)
2066
2067
2068
2069
2070
2071
2072
2073

    def test_torch_compile_repeated_blocks(self):
        if self.model_class._repeated_blocks is None:
            pytest.skip("Skipping test as the model class doesn't have `_repeated_blocks` set.")

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict).to(torch_device)
2074
        model.eval()
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
        model.compile_repeated_blocks(fullgraph=True)

        recompile_limit = 1
        if self.model_class.__name__ == "UNet2DConditionModel":
            recompile_limit = 2

        with (
            torch._inductor.utils.fresh_inductor_cache(),
            torch._dynamo.config.patch(recompile_limit=recompile_limit),
            torch.no_grad(),
        ):
            _ = model(**inputs_dict)
            _ = model(**inputs_dict)
2088

2089
    def test_compile_with_group_offloading(self):
2090
2091
2092
        if not self.model_class._supports_group_offloading:
            pytest.skip("Model does not support group offloading.")

2093
2094
2095
2096
2097
2098
2099
        torch._dynamo.config.cache_size_limit = 10000

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)
        model.eval()
        # TODO: Can test for other group offloading kwargs later if needed.
        group_offload_kwargs = {
2100
            "onload_device": torch_device,
2101
2102
2103
2104
2105
2106
2107
2108
            "offload_device": "cpu",
            "offload_type": "block_level",
            "num_blocks_per_group": 1,
            "use_stream": True,
            "non_blocking": True,
        }
        model.enable_group_offload(**group_offload_kwargs)
        model.compile()
2109

2110
2111
2112
2113
        with torch.no_grad():
            _ = model(**inputs_dict)
            _ = model(**inputs_dict)

2114
2115
2116
2117
2118
2119
2120
    def test_compile_on_different_shapes(self):
        if self.different_shapes_for_compilation is None:
            pytest.skip(f"Skipping as `different_shapes_for_compilation` is not set for {self.__class__.__name__}.")
        torch.fx.experimental._config.use_duck_shape = False

        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)
2121
        model.eval()
2122
2123
2124
2125
2126
2127
2128
        model = torch.compile(model, fullgraph=True, dynamic=True)

        for height, width in self.different_shapes_for_compilation:
            with torch._dynamo.config.patch(error_on_recompile=True), torch.no_grad():
                inputs_dict = self.prepare_dummy_input(height=height, width=width)
                _ = model(**inputs_dict)

2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
    def test_compile_works_with_aot(self):
        from torch._inductor.package import load_package

        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()

        model = self.model_class(**init_dict).to(torch_device)
        exported_model = torch.export.export(model, args=(), kwargs=inputs_dict)

        with tempfile.TemporaryDirectory() as tmpdir:
            package_path = os.path.join(tmpdir, f"{self.model_class.__name__}.pt2")
            _ = torch._inductor.aoti_compile_and_package(exported_model, package_path=package_path)
            assert os.path.exists(package_path)
            loaded_binary = load_package(package_path, run_single_threaded=True)

        model.forward = loaded_binary

        with torch.no_grad():
            _ = model(**inputs_dict)
            _ = model(**inputs_dict)

2149

2150
2151
2152
2153
2154
@slow
@require_torch_2
@require_torch_accelerator
@require_peft_backend
@require_peft_version_greater("0.14.0")
2155
@require_torch_version_greater("2.7.1")
2156
@is_torch_compile
2157
class LoraHotSwappingForModelTesterMixin:
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
    """Test that hotswapping does not result in recompilation on the model directly.

    We're not extensively testing the hotswapping functionality since it is implemented in PEFT and is extensively
    tested there. The goal of this test is specifically to ensure that hotswapping with diffusers does not require
    recompilation.

    See
    https://github.com/huggingface/peft/blob/eaab05e18d51fb4cce20a73c9acd82a00c013b83/tests/test_gpu_examples.py#L4252
    for the analogous PEFT test.

    """

2170
2171
    different_shapes_for_compilation = None

2172
2173
2174
2175
    def tearDown(self):
        # It is critical that the dynamo cache is reset for each test. Otherwise, if the test re-uses the same model,
        # there will be recompilation errors, as torch caches the model when run in the same process.
        super().tearDown()
2176
        torch.compiler.reset()
2177
2178
2179
        gc.collect()
        backend_empty_cache(torch_device)

2180
    def get_lora_config(self, lora_rank, lora_alpha, target_modules):
2181
2182
2183
        # from diffusers test_models_unet_2d_condition.py
        from peft import LoraConfig

2184
        lora_config = LoraConfig(
2185
2186
2187
2188
2189
2190
            r=lora_rank,
            lora_alpha=lora_alpha,
            target_modules=target_modules,
            init_lora_weights=False,
            use_dora=False,
        )
2191
        return lora_config
2192

2193
2194
2195
2196
2197
    def get_linear_module_name_other_than_attn(self, model):
        linear_names = [
            name for name, module in model.named_modules() if isinstance(module, nn.Linear) and "to_" not in name
        ]
        return linear_names[0]
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208

    def check_model_hotswap(self, do_compile, rank0, rank1, target_modules0, target_modules1=None):
        """
        Check that hotswapping works on a small unet.

        Steps:
        - create 2 LoRA adapters and save them
        - load the first adapter
        - hotswap the second adapter
        - check that the outputs are correct
        - optionally compile the model
2209
        - optionally check if recompilations happen on different shapes
2210
2211
2212
2213
2214

        Note: We set rank == alpha here because save_lora_adapter does not save the alpha scalings, thus the test would
        fail if the values are different. Since rank != alpha does not matter for the purpose of this test, this is
        fine.
        """
2215
        different_shapes = self.different_shapes_for_compilation
2216
        # create 2 adapters with different ranks and alphas
2217
2218
2219
2220
        torch.manual_seed(0)
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)

2221
2222
2223
2224
        alpha0, alpha1 = rank0, rank1
        max_rank = max([rank0, rank1])
        if target_modules1 is None:
            target_modules1 = target_modules0[:]
2225
2226
        lora_config0 = self.get_lora_config(rank0, alpha0, target_modules0)
        lora_config1 = self.get_lora_config(rank1, alpha1, target_modules1)
2227

2228
        model.add_adapter(lora_config0, adapter_name="adapter0")
2229
        with torch.inference_mode():
2230
2231
            torch.manual_seed(0)
            output0_before = model(**inputs_dict)["sample"]
2232

2233
2234
        model.add_adapter(lora_config1, adapter_name="adapter1")
        model.set_adapter("adapter1")
2235
        with torch.inference_mode():
2236
2237
            torch.manual_seed(0)
            output1_before = model(**inputs_dict)["sample"]
2238
2239
2240
2241
2242
2243
2244
2245
2246

        # sanity checks:
        tol = 5e-3
        assert not torch.allclose(output0_before, output1_before, atol=tol, rtol=tol)
        assert not (output0_before == 0).all()
        assert not (output1_before == 0).all()

        with tempfile.TemporaryDirectory() as tmp_dirname:
            # save the adapter checkpoints
2247
2248
2249
            model.save_lora_adapter(os.path.join(tmp_dirname, "0"), safe_serialization=True, adapter_name="adapter0")
            model.save_lora_adapter(os.path.join(tmp_dirname, "1"), safe_serialization=True, adapter_name="adapter1")
            del model
2250
2251

            # load the first adapter
2252
2253
2254
2255
            torch.manual_seed(0)
            init_dict, _ = self.prepare_init_args_and_inputs_for_common()
            model = self.model_class(**init_dict).to(torch_device)

2256
2257
            if do_compile or (rank0 != rank1):
                # no need to prepare if the model is not compiled or if the ranks are identical
2258
                model.enable_lora_hotswap(target_rank=max_rank)
2259
2260
2261

            file_name0 = os.path.join(os.path.join(tmp_dirname, "0"), "pytorch_lora_weights.safetensors")
            file_name1 = os.path.join(os.path.join(tmp_dirname, "1"), "pytorch_lora_weights.safetensors")
2262
            model.load_lora_adapter(file_name0, safe_serialization=True, adapter_name="adapter0", prefix=None)
2263
2264

            if do_compile:
2265
                model = torch.compile(model, mode="reduce-overhead", dynamic=different_shapes is not None)
2266
2267

            with torch.inference_mode():
2268
2269
2270
2271
2272
2273
2274
2275
                # additionally check if dynamic compilation works.
                if different_shapes is not None:
                    for height, width in different_shapes:
                        new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
                        _ = model(**new_inputs_dict)
                else:
                    output0_after = model(**inputs_dict)["sample"]
                    assert torch.allclose(output0_before, output0_after, atol=tol, rtol=tol)
2276
2277

            # hotswap the 2nd adapter
2278
            model.load_lora_adapter(file_name1, adapter_name="adapter0", hotswap=True, prefix=None)
2279
2280
2281

            # we need to call forward to potentially trigger recompilation
            with torch.inference_mode():
2282
2283
2284
2285
2286
2287
2288
                if different_shapes is not None:
                    for height, width in different_shapes:
                        new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
                        _ = model(**new_inputs_dict)
                else:
                    output1_after = model(**inputs_dict)["sample"]
                    assert torch.allclose(output1_before, output1_after, atol=tol, rtol=tol)
2289
2290
2291
2292
2293

            # check error when not passing valid adapter name
            name = "does-not-exist"
            msg = f"Trying to hotswap LoRA adapter '{name}' but there is no existing adapter by that name"
            with self.assertRaisesRegex(ValueError, msg):
2294
                model.load_lora_adapter(file_name1, adapter_name=name, hotswap=True, prefix=None)
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305

    @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
    def test_hotswapping_model(self, rank0, rank1):
        self.check_model_hotswap(
            do_compile=False, rank0=rank0, rank1=rank1, target_modules0=["to_q", "to_k", "to_v", "to_out.0"]
        )

    @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
    def test_hotswapping_compiled_model_linear(self, rank0, rank1):
        # It's important to add this context to raise an error on recompilation
        target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
2306
        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
2307
2308
2309
2310
            self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)

    @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
    def test_hotswapping_compiled_model_conv2d(self, rank0, rank1):
2311
        if "unet" not in self.model_class.__name__.lower():
2312
            pytest.skip("Test only applies to UNet.")
2313

2314
2315
        # It's important to add this context to raise an error on recompilation
        target_modules = ["conv", "conv1", "conv2"]
2316
        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
2317
2318
2319
2320
            self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)

    @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
    def test_hotswapping_compiled_model_both_linear_and_conv2d(self, rank0, rank1):
2321
        if "unet" not in self.model_class.__name__.lower():
2322
            pytest.skip("Test only applies to UNet.")
2323

2324
2325
        # It's important to add this context to raise an error on recompilation
        target_modules = ["to_q", "conv"]
2326
        with torch._dynamo.config.patch(error_on_recompile=True), torch._inductor.utils.fresh_inductor_cache():
2327
2328
            self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)

2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
    @parameterized.expand([(11, 11), (7, 13), (13, 7)])  # important to test small to large and vice versa
    def test_hotswapping_compiled_model_both_linear_and_other(self, rank0, rank1):
        # In `test_hotswapping_compiled_model_both_linear_and_conv2d()`, we check if we can do hotswapping
        # with `torch.compile()` for models that have both linear and conv layers. In this test, we check
        # if we can target a linear layer from the transformer blocks and another linear layer from non-attention
        # block.
        target_modules = ["to_q"]
        init_dict, _ = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict)

        target_modules.append(self.get_linear_module_name_other_than_attn(model))
        del model

        # It's important to add this context to raise an error on recompilation
        with torch._dynamo.config.patch(error_on_recompile=True):
            self.check_model_hotswap(do_compile=True, rank0=rank0, rank1=rank1, target_modules0=target_modules)

2346
2347
    def test_enable_lora_hotswap_called_after_adapter_added_raises(self):
        # ensure that enable_lora_hotswap is called before loading the first adapter
2348
2349
2350
2351
2352
        lora_config = self.get_lora_config(8, 8, target_modules=["to_q"])
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)
        model.add_adapter(lora_config)

2353
2354
        msg = re.escape("Call `enable_lora_hotswap` before loading the first adapter.")
        with self.assertRaisesRegex(RuntimeError, msg):
2355
            model.enable_lora_hotswap(target_rank=32)
2356
2357
2358
2359
2360

    def test_enable_lora_hotswap_called_after_adapter_added_warning(self):
        # ensure that enable_lora_hotswap is called before loading the first adapter
        from diffusers.loaders.peft import logger

2361
2362
2363
2364
        lora_config = self.get_lora_config(8, 8, target_modules=["to_q"])
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)
        model.add_adapter(lora_config)
2365
2366
2367
2368
        msg = (
            "It is recommended to call `enable_lora_hotswap` before loading the first adapter to avoid recompilation."
        )
        with self.assertLogs(logger=logger, level="WARNING") as cm:
2369
            model.enable_lora_hotswap(target_rank=32, check_compiled="warn")
2370
2371
2372
2373
            assert any(msg in log for log in cm.output)

    def test_enable_lora_hotswap_called_after_adapter_added_ignore(self):
        # check possibility to ignore the error/warning
2374
2375
2376
2377
        lora_config = self.get_lora_config(8, 8, target_modules=["to_q"])
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)
        model.add_adapter(lora_config)
2378
2379
        with warnings.catch_warnings(record=True) as w:
            warnings.simplefilter("always")  # Capture all warnings
2380
            model.enable_lora_hotswap(target_rank=32, check_compiled="warn")
2381
2382
2383
2384
            self.assertEqual(len(w), 0, f"Expected no warnings, but got: {[str(warn.message) for warn in w]}")

    def test_enable_lora_hotswap_wrong_check_compiled_argument_raises(self):
        # check that wrong argument value raises an error
2385
2386
2387
2388
        lora_config = self.get_lora_config(8, 8, target_modules=["to_q"])
        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
        model = self.model_class(**init_dict).to(torch_device)
        model.add_adapter(lora_config)
2389
2390
        msg = re.escape("check_compiles should be one of 'error', 'warn', or 'ignore', got 'wrong-argument' instead.")
        with self.assertRaisesRegex(ValueError, msg):
2391
            model.enable_lora_hotswap(target_rank=32, check_compiled="wrong-argument")
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405

    def test_hotswap_second_adapter_targets_more_layers_raises(self):
        # check the error and log
        from diffusers.loaders.peft import logger

        # at the moment, PEFT requires the 2nd adapter to target the same or a subset of layers
        target_modules0 = ["to_q"]
        target_modules1 = ["to_q", "to_k"]
        with self.assertRaises(RuntimeError):  # peft raises RuntimeError
            with self.assertLogs(logger=logger, level="ERROR") as cm:
                self.check_model_hotswap(
                    do_compile=True, rank0=8, rank1=8, target_modules0=target_modules0, target_modules1=target_modules1
                )
                assert any("Hotswapping adapter0 was unsuccessful" in log for log in cm.output)
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425

    @parameterized.expand([(11, 11), (7, 13), (13, 7)])
    @require_torch_version_greater("2.7.1")
    def test_hotswapping_compile_on_different_shapes(self, rank0, rank1):
        different_shapes_for_compilation = self.different_shapes_for_compilation
        if different_shapes_for_compilation is None:
            pytest.skip(f"Skipping as `different_shapes_for_compilation` is not set for {self.__class__.__name__}.")
        # Specifying `use_duck_shape=False` instructs the compiler if it should use the same symbolic
        # variable to represent input sizes that are the same. For more details,
        # check out this [comment](https://github.com/huggingface/diffusers/pull/11327#discussion_r2047659790).
        torch.fx.experimental._config.use_duck_shape = False

        target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
        with torch._dynamo.config.patch(error_on_recompile=True):
            self.check_model_hotswap(
                do_compile=True,
                rank0=rank0,
                rank1=rank1,
                target_modules0=target_modules,
            )