{{ message }}
Repository navigation
Expand file tree
/
Copy pathtest_group_offloading.py
More file actions
696 lines (556 loc) · 29.7 KB
/
Copy pathtest_group_offloading.py
File metadata and controls
696 lines (556 loc) · 29.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
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
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
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
# Copyright 2026 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import contextlib
import gc
import logging
import pytest
import torch
from diffusers import AutoencoderKL
from diffusers.hooks import HookRegistry, ModelHook
from diffusers.models import ModelMixin
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.utils import logging as diffusers_logging
from diffusers.utils.import_utils import compare_versions
@contextlib.contextmanager
def _propagate_diffusers_logs():
# The diffusers root logger sets `propagate = False`, so its records never reach the root
# logger that pytest's `caplog` fixture hooks into. Temporarily enable propagation so the
# emitted warnings can be captured, then restore the default.
diffusers_logging.enable_propagation()
try:
yield
finally:
diffusers_logging.disable_propagation()
from ..testing_utils import ( # noqa: E402
backend_empty_cache,
backend_max_memory_allocated,
backend_reset_peak_memory_stats,
require_torch_accelerator,
torch_device,
)
class DummyBlock(torch.nn.Module):
def __init__(self, in_features: int, hidden_features: int, out_features: int) -> None:
super().__init__()
self.proj_in = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.proj_out = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.proj_in(x)
x = self.activation(x)
x = self.proj_out(x)
return x
class DummyModel(ModelMixin):
def __init__(self, in_features: int, hidden_features: int, out_features: int, num_layers: int) -> None:
super().__init__()
self.linear_1 = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.blocks = torch.nn.ModuleList(
[DummyBlock(hidden_features, hidden_features, hidden_features) for _ in range(num_layers)]
)
self.linear_2 = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.linear_1(x)
x = self.activation(x)
for block in self.blocks:
x = block(x)
x = self.linear_2(x)
return x
# This model implementation contains one type of block (single_blocks) instantiated before another type of block (double_blocks).
# The invocation order of these blocks, however, is first the double_blocks and then the single_blocks.
# With group offloading implementation before https://github.com/huggingface/diffusers/pull/11375, such a modeling implementation
# would result in a device mismatch error because of the assumptions made by the code. The failure case occurs when using:
# offload_type="block_level", num_blocks_per_group=2, use_stream=True
# Post the linked PR, the implementation will work as expected.
class DummyModelWithMultipleBlocks(ModelMixin):
def __init__(
self, in_features: int, hidden_features: int, out_features: int, num_layers: int, num_single_layers: int
) -> None:
super().__init__()
self.linear_1 = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.single_blocks = torch.nn.ModuleList(
[DummyBlock(hidden_features, hidden_features, hidden_features) for _ in range(num_single_layers)]
)
self.double_blocks = torch.nn.ModuleList(
[DummyBlock(hidden_features, hidden_features, hidden_features) for _ in range(num_layers)]
)
self.linear_2 = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.linear_1(x)
x = self.activation(x)
for block in self.double_blocks:
x = block(x)
for block in self.single_blocks:
x = block(x)
x = self.linear_2(x)
return x
# Test for https://github.com/huggingface/diffusers/pull/12077
class DummyModelWithLayerNorm(ModelMixin):
def __init__(self, in_features: int, hidden_features: int, out_features: int, num_layers: int) -> None:
super().__init__()
self.linear_1 = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.blocks = torch.nn.ModuleList(
[DummyBlock(hidden_features, hidden_features, hidden_features) for _ in range(num_layers)]
)
self.layer_norm = torch.nn.LayerNorm(hidden_features, elementwise_affine=True)
self.linear_2 = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.linear_1(x)
x = self.activation(x)
for block in self.blocks:
x = block(x)
x = self.layer_norm(x)
x = self.linear_2(x)
return x
class DummyPipeline(DiffusionPipeline):
model_cpu_offload_seq = "model"
def __init__(self, model: torch.nn.Module) -> None:
super().__init__()
self.register_modules(model=model)
def __call__(self, x: torch.Tensor) -> torch.Tensor:
for _ in range(2):
x = x + 0.1 * self.model(x)
return x
class LayerOutputTrackerHook(ModelHook):
def __init__(self):
super().__init__()
self.outputs = []
def post_forward(self, module, output):
self.outputs.append(output)
return output
# Model with only standalone computational layers at top level
class DummyModelWithStandaloneLayers(ModelMixin):
def __init__(self, in_features: int, hidden_features: int, out_features: int) -> None:
super().__init__()
self.layer1 = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.layer2 = torch.nn.Linear(hidden_features, hidden_features)
self.layer3 = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.layer1(x)
x = self.activation(x)
x = self.layer2(x)
x = self.layer3(x)
return x
# Model with deeply nested structure
class DummyModelWithDeeplyNestedBlocks(ModelMixin):
def __init__(self, in_features: int, hidden_features: int, out_features: int) -> None:
super().__init__()
self.input_layer = torch.nn.Linear(in_features, hidden_features)
self.container = ContainerWithNestedModuleList(hidden_features)
self.output_layer = torch.nn.Linear(hidden_features, out_features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.input_layer(x)
x = self.container(x)
x = self.output_layer(x)
return x
class ContainerWithNestedModuleList(torch.nn.Module):
def __init__(self, features: int) -> None:
super().__init__()
# Top-level computational layer
self.proj_in = torch.nn.Linear(features, features)
# Nested container with ModuleList
self.nested_container = NestedContainer(features)
# Another top-level computational layer
self.proj_out = torch.nn.Linear(features, features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.proj_in(x)
x = self.nested_container(x)
x = self.proj_out(x)
return x
class NestedContainer(torch.nn.Module):
def __init__(self, features: int) -> None:
super().__init__()
self.blocks = torch.nn.ModuleList([torch.nn.Linear(features, features), torch.nn.Linear(features, features)])
self.norm = torch.nn.LayerNorm(features)
def forward(self, x: torch.Tensor) -> torch.Tensor:
for block in self.blocks:
x = block(x)
x = self.norm(x)
return x
@require_torch_accelerator
class TestGroupOffload:
in_features = 64
hidden_features = 256
out_features = 64
num_layers = 4
def setup_method(self):
with torch.no_grad():
self.model = self.get_model()
self.input = torch.randn((4, self.in_features)).to(torch_device)
def teardown_method(self):
del self.model
del self.input
gc.collect()
backend_empty_cache(torch_device)
backend_reset_peak_memory_stats(torch_device)
def get_model(self):
torch.manual_seed(0)
return DummyModel(
in_features=self.in_features,
hidden_features=self.hidden_features,
out_features=self.out_features,
num_layers=self.num_layers,
)
@pytest.mark.skipif(
torch.device(torch_device).type not in ["cuda", "xpu"],
reason="Test requires a CUDA or XPU device.",
)
def test_offloading_forward_pass(self):
@torch.no_grad()
def run_forward(model):
gc.collect()
backend_empty_cache(torch_device)
backend_reset_peak_memory_stats(torch_device)
assert all(
module._diffusers_hook.get_hook("group_offloading") is not None
for module in model.modules()
if hasattr(module, "_diffusers_hook")
)
model.eval()
output = model(self.input)[0].cpu()
max_memory_allocated = backend_max_memory_allocated(torch_device)
return output, max_memory_allocated
self.model.to(torch_device)
output_without_group_offloading, mem_baseline = run_forward(self.model)
self.model.to("cpu")
model = self.get_model()
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
output_with_group_offloading1, mem1 = run_forward(model)
model = self.get_model()
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1)
output_with_group_offloading2, mem2 = run_forward(model)
model = self.get_model()
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=True)
output_with_group_offloading3, mem3 = run_forward(model)
model = self.get_model()
model.enable_group_offload(torch_device, offload_type="leaf_level")
output_with_group_offloading4, mem4 = run_forward(model)
model = self.get_model()
model.enable_group_offload(torch_device, offload_type="leaf_level", use_stream=True)
output_with_group_offloading5, mem5 = run_forward(model)
# Precision assertions - offloading should not impact the output
assert torch.allclose(output_without_group_offloading, output_with_group_offloading1, atol=1e-5)
assert torch.allclose(output_without_group_offloading, output_with_group_offloading2, atol=1e-5)
assert torch.allclose(output_without_group_offloading, output_with_group_offloading3, atol=1e-5)
assert torch.allclose(output_without_group_offloading, output_with_group_offloading4, atol=1e-5)
assert torch.allclose(output_without_group_offloading, output_with_group_offloading5, atol=1e-5)
# Memory assertions - offloading should reduce memory usage
assert mem4 <= mem5 < mem2 <= mem3 < mem1 < mem_baseline
def test_warning_logged_if_group_offloaded_module_moved_to_accelerator(self, caplog):
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
self.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
with _propagate_diffusers_logs(), caplog.at_level(logging.WARNING, logger="diffusers.models.modeling_utils"):
self.model.to(torch_device)
assert f"The module '{self.model.__class__.__name__}' is group offloaded" in caplog.text
def test_warning_logged_if_group_offloaded_pipe_moved_to_accelerator(self, caplog):
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
pipe = DummyPipeline(self.model)
self.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
with (
_propagate_diffusers_logs(),
caplog.at_level(logging.WARNING, logger="diffusers.pipelines.pipeline_utils"),
):
pipe.to(torch_device)
assert f"The module '{self.model.__class__.__name__}' is group offloaded" in caplog.text
def test_error_raised_if_streams_used_and_no_accelerator_device(self):
with pytest.raises(ValueError, match="backend implements streams, got `cpu`"):
self.model.enable_group_offload(
onload_device=torch.device("cpu"), offload_type="leaf_level", use_stream=True
)
def test_error_raised_if_streams_used_and_backend_has_no_streams(self):
with pytest.raises(ValueError, match="backend implements streams, got `mps`"):
self.model.enable_group_offload(
onload_device=torch.device("mps"), offload_type="leaf_level", use_stream=True
)
def test_error_raised_if_supports_group_offloading_false(self):
self.model._supports_group_offloading = False
with pytest.raises(ValueError, match="does not support group offloading"):
self.model.enable_group_offload(onload_device=torch.device(torch_device))
def test_error_raised_if_model_offloading_applied_on_group_offloaded_module(self):
pipe = DummyPipeline(self.model)
pipe.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
with pytest.raises(ValueError, match="You are trying to apply model/sequential CPU offloading"):
pipe.enable_model_cpu_offload()
def test_error_raised_if_sequential_offloading_applied_on_group_offloaded_module(self):
pipe = DummyPipeline(self.model)
pipe.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
with pytest.raises(ValueError, match="You are trying to apply model/sequential CPU offloading"):
pipe.enable_sequential_cpu_offload()
def test_error_raised_if_group_offloading_applied_on_model_offloaded_module(self):
pipe = DummyPipeline(self.model)
pipe.enable_model_cpu_offload()
with pytest.raises(ValueError, match="Cannot apply group offloading"):
pipe.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
def test_error_raised_if_group_offloading_applied_on_sequential_offloaded_module(self):
pipe = DummyPipeline(self.model)
pipe.enable_sequential_cpu_offload()
with pytest.raises(ValueError, match="Cannot apply group offloading"):
pipe.model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=3)
def test_block_level_stream_with_invocation_order_different_from_initialization_order(self):
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
model = DummyModelWithMultipleBlocks(
in_features=self.in_features,
hidden_features=self.hidden_features,
out_features=self.out_features,
num_layers=self.num_layers,
num_single_layers=self.num_layers + 1,
)
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=True)
context = contextlib.nullcontext()
if compare_versions("diffusers", "<=", "0.33.0"):
# Will raise a device mismatch RuntimeError mentioning weights are on CPU but input is on device
context = pytest.raises(RuntimeError, match="Expected all tensors to be on the same device")
with context:
model(self.input)
@pytest.mark.parametrize("offload_type", ["block_level", "leaf_level"])
def test_block_level_offloading_with_parameter_only_module_group(self, offload_type: str):
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
def apply_layer_output_tracker_hook(model: DummyModelWithLayerNorm):
for name, module in model.named_modules():
registry = HookRegistry.check_if_exists_or_initialize(module)
hook = LayerOutputTrackerHook()
registry.register_hook(hook, "layer_output_tracker")
model_ref = DummyModelWithLayerNorm(128, 256, 128, 2)
model = DummyModelWithLayerNorm(128, 256, 128, 2)
model.load_state_dict(model_ref.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type=offload_type, num_blocks_per_group=1, use_stream=True)
apply_layer_output_tracker_hook(model_ref)
apply_layer_output_tracker_hook(model)
x = torch.randn(2, 128).to(torch_device)
out_ref = model_ref(x)
out = model(x)
assert torch.allclose(out_ref, out, atol=1e-5), "Outputs do not match."
num_repeats = 2
for i in range(num_repeats):
out_ref = model_ref(x)
out = model(x)
assert torch.allclose(out_ref, out, atol=1e-5), "Outputs do not match after multiple invocations."
for (ref_name, ref_module), (name, module) in zip(model_ref.named_modules(), model.named_modules()):
assert ref_name == name
ref_outputs = (
HookRegistry.check_if_exists_or_initialize(ref_module).get_hook("layer_output_tracker").outputs
)
outputs = HookRegistry.check_if_exists_or_initialize(module).get_hook("layer_output_tracker").outputs
cumulated_absmax = 0.0
for i in range(len(outputs)):
diff = ref_outputs[0] - outputs[i]
absdiff = diff.abs()
absmax = absdiff.max().item()
cumulated_absmax += absmax
assert cumulated_absmax < 1e-5, f"Output differences for {name} exceeded threshold: {cumulated_absmax:.5f}"
def test_vae_like_model_without_streams(self):
"""Test VAE-like model with block-level offloading but without streams."""
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
config = self.get_autoencoder_kl_config()
model = AutoencoderKL(**config)
model_ref = AutoencoderKL(**config)
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=False)
x = torch.randn(2, 3, 32, 32).to(torch_device)
with torch.no_grad():
out_ref = model_ref(x).sample
out = model(x).sample
assert torch.allclose(out_ref, out, atol=1e-5), "Outputs do not match for VAE-like model without streams."
def test_model_with_only_standalone_layers(self):
"""Test that models with only standalone layers (no ModuleList/Sequential) work with block-level offloading."""
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
model = DummyModelWithStandaloneLayers(in_features=64, hidden_features=128, out_features=64)
model_ref = DummyModelWithStandaloneLayers(in_features=64, hidden_features=128, out_features=64)
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=True)
x = torch.randn(2, 64).to(torch_device)
with torch.no_grad():
for i in range(2):
out_ref = model_ref(x)
out = model(x)
assert torch.allclose(out_ref, out, atol=1e-5), (
f"Outputs do not match at iteration {i} for model with standalone layers."
)
@pytest.mark.parametrize("offload_type", ["block_level", "leaf_level"])
def test_standalone_conv_layers_with_both_offload_types(self, offload_type: str):
"""Test that standalone Conv2d layers work correctly with both block-level and leaf-level offloading."""
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
config = self.get_autoencoder_kl_config()
model = AutoencoderKL(**config)
model_ref = AutoencoderKL(**config)
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type=offload_type, num_blocks_per_group=1, use_stream=True)
x = torch.randn(2, 3, 32, 32).to(torch_device)
with torch.no_grad():
out_ref = model_ref(x).sample
out = model(x).sample
assert torch.allclose(out_ref, out, atol=1e-5), (
f"Outputs do not match for standalone Conv layers with {offload_type}."
)
def test_multiple_invocations_with_vae_like_model(self):
"""Test that multiple forward passes work correctly with VAE-like model."""
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
config = self.get_autoencoder_kl_config()
model = AutoencoderKL(**config)
model_ref = AutoencoderKL(**config)
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=True)
x = torch.randn(2, 3, 32, 32).to(torch_device)
with torch.no_grad():
for i in range(2):
out_ref = model_ref(x).sample
out = model(x).sample
assert torch.allclose(out_ref, out, atol=1e-5), f"Outputs do not match at iteration {i}."
def test_nested_container_parameters_offloading(self):
"""Test that parameters from non-computational layers in nested containers are handled correctly."""
if torch.device(torch_device).type not in ["cuda", "xpu"]:
return
model = DummyModelWithDeeplyNestedBlocks(in_features=64, hidden_features=128, out_features=64)
model_ref = DummyModelWithDeeplyNestedBlocks(in_features=64, hidden_features=128, out_features=64)
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(torch_device, offload_type="block_level", num_blocks_per_group=1, use_stream=True)
x = torch.randn(2, 64).to(torch_device)
with torch.no_grad():
for i in range(2):
out_ref = model_ref(x)
out = model(x)
assert torch.allclose(out_ref, out, atol=1e-5), (
f"Outputs do not match at iteration {i} for nested parameters."
)
def get_autoencoder_kl_config(self, block_out_channels=None, norm_num_groups=None):
block_out_channels = block_out_channels or [2, 4]
norm_num_groups = norm_num_groups or 2
init_dict = {
"block_out_channels": block_out_channels,
"in_channels": 3,
"out_channels": 3,
"down_block_types": ["DownEncoderBlock2D"] * len(block_out_channels),
"up_block_types": ["UpDecoderBlock2D"] * len(block_out_channels),
"latent_channels": 4,
"norm_num_groups": norm_num_groups,
"layers_per_block": 1,
}
return init_dict
# Model with conditionally-executed modules, simulating Helios patch_short/patch_mid/patch_long behavior.
# These modules are only called when optional inputs are provided, which means the lazy prefetch
# execution order tracer may not see them on the first forward pass. This can cause a device mismatch
# on subsequent calls when the modules ARE invoked but their weights were never onloaded.
# See: https://github.com/huggingface/diffusers/pull/13211
class DummyModelWithConditionalModules(ModelMixin):
def __init__(self, in_features: int, hidden_features: int, out_features: int, num_layers: int) -> None:
super().__init__()
self.linear_1 = torch.nn.Linear(in_features, hidden_features)
self.activation = torch.nn.ReLU()
self.blocks = torch.nn.ModuleList(
[DummyBlock(hidden_features, hidden_features, hidden_features) for _ in range(num_layers)]
)
self.linear_2 = torch.nn.Linear(hidden_features, out_features)
# These modules are only invoked when optional_input is not None.
# Output dimension matches hidden_features so they can be added after linear_1.
self.optional_proj_1 = torch.nn.Linear(in_features, hidden_features)
self.optional_proj_2 = torch.nn.Linear(in_features, hidden_features)
def forward(self, x: torch.Tensor, optional_input: torch.Tensor | None = None) -> torch.Tensor:
x = self.linear_1(x)
x = self.activation(x)
if optional_input is not None:
# Add optional projections after linear_1 so dimensions match (both hidden_features)
x = x + self.optional_proj_1(optional_input)
x = x + self.optional_proj_2(optional_input)
for block in self.blocks:
x = block(x)
x = self.linear_2(x)
return x
class TestConditionalModuleGroupOffload(TestGroupOffload):
"""Tests for conditionally-executed modules under group offloading with streams.
Regression tests for the case where a module is not executed during the first forward pass
(when the lazy prefetch execution order is traced), but IS executed on subsequent passes.
Without the fix, the weights of such modules remain on CPU while the input is on GPU,
causing a RuntimeError about tensor device mismatch.
"""
def get_model(self):
torch.manual_seed(0)
return DummyModelWithConditionalModules(
in_features=self.in_features,
hidden_features=self.hidden_features,
out_features=self.out_features,
num_layers=self.num_layers,
)
@pytest.mark.parametrize("offload_type", ["leaf_level", "block_level"])
@pytest.mark.skipif(
torch.device(torch_device).type not in ["cuda", "xpu"],
reason="Test requires a CUDA or XPU device.",
)
def test_conditional_modules_with_stream(self, offload_type: str):
"""Regression test: conditionally-executed modules must not cause device mismatch when using streams.
The model contains two optional Linear layers (optional_proj_1, optional_proj_2) that are only
executed when `optional_input` is provided. This simulates modules like patch_short/patch_mid/
patch_long in HeliosTransformer3DModel, which are only called when history latents are present.
When using streams, `LazyPrefetchGroupOffloadingHook` traces the execution order on the first
forward pass and sets up a prefetch chain so each module pre-loads the next one's weights.
Modules not executed during this tracing pass are excluded from the prefetch chain.
The bug: if a module was absent from the first (tracing) pass, its `onload_self` flag gets set
to False (meaning "someone else will onload me"). But since it's not in the prefetch chain,
nobody ever does — so its weights remain on CPU. When the module is eventually called in a
subsequent pass, the input is on GPU but the weights are on CPU, causing a RuntimeError.
We therefore must invoke the model multiple times:
1. First pass WITHOUT optional_input: triggers the lazy prefetch tracing. optional_proj_1/2
are absent, so they are excluded from the prefetch chain.
2. Second pass WITH optional_input: the regression case. Without the fix, this raises a
RuntimeError because optional_proj_1/2 weights are still on CPU.
3. Third pass WITHOUT optional_input: verifies the model remains stable after having seen
both code paths.
"""
model = self.get_model()
model_ref = self.get_model()
model_ref.load_state_dict(model.state_dict(), strict=True)
model_ref.to(torch_device)
model.enable_group_offload(
torch_device,
offload_type=offload_type,
num_blocks_per_group=1,
use_stream=True,
)
x = torch.randn(4, self.in_features).to(torch_device)
optional_input = torch.randn(4, self.in_features).to(torch_device)
with torch.no_grad():
# First forward pass WITHOUT optional_input — this is when the lazy prefetch
# execution order is traced. optional_proj_1/2 are NOT in the traced order.
out_ref_no_opt = model_ref(x, optional_input=None)
out_no_opt = model(x, optional_input=None)
assert torch.allclose(out_ref_no_opt, out_no_opt, atol=1e-5), (
f"[{offload_type}] Outputs do not match on first pass (no optional_input)."
)
# Second forward pass WITH optional_input — optional_proj_1/2 ARE now called.
out_ref_with_opt = model_ref(x, optional_input=optional_input)
out_with_opt = model(x, optional_input=optional_input)
assert torch.allclose(out_ref_with_opt, out_with_opt, atol=1e-5), (
f"[{offload_type}] Outputs do not match on second pass (with optional_input)."
)
# Third pass again without optional_input — verify stable behavior.
out_ref_no_opt2 = model_ref(x, optional_input=None)
out_no_opt2 = model(x, optional_input=None)
assert torch.allclose(out_ref_no_opt2, out_no_opt2, atol=1e-5), (
f"[{offload_type}] Outputs do not match on third pass (back to no optional_input)."
)
You can’t perform that action at this time.