forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathgroup_offloading.py
More file actions
1056 lines (891 loc) 路 49.7 KB
/
Copy pathgroup_offloading.py
File metadata and controls
1056 lines (891 loc) 路 49.7 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
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2026 The HuggingFace Team. All rights reserved.
#
# 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 hashlib
import os
from contextlib import contextmanager, nullcontext
from dataclasses import dataclass, replace
from enum import Enum
from typing import Set
import safetensors.torch
import torch
from ..utils import get_logger, is_accelerate_available, is_torchao_available
from ._common import _GO_LC_SUPPORTED_PYTORCH_LAYERS
from .hooks import HookRegistry, ModelHook
if is_accelerate_available():
from accelerate.hooks import AlignDevicesHook, CpuOffload
from accelerate.utils import send_to_device
logger = get_logger(__name__) # pylint: disable=invalid-name
def _is_torchao_tensor(tensor: torch.Tensor) -> bool:
if not is_torchao_available():
return False
from torchao.utils import TorchAOBaseTensor
return isinstance(tensor, TorchAOBaseTensor)
def _get_torchao_inner_tensor_names(tensor: torch.Tensor) -> list[str]:
"""Get names of all internal tensor data attributes from a TorchAO tensor."""
cls = type(tensor)
names = list(getattr(cls, "tensor_data_names", []))
for attr_name in getattr(cls, "optional_tensor_data_names", []):
if getattr(tensor, attr_name, None) is not None:
names.append(attr_name)
return names
def _swap_torchao_tensor(param: torch.Tensor, source: torch.Tensor) -> None:
"""Move a TorchAO parameter to the device of `source` via `swap_tensors`.
`param.data = source` does not work for `_make_wrapper_subclass` tensors because the `.data` setter only replaces
the outer wrapper storage while leaving the subclass's internal attributes (e.g. `.qdata`, `.scale`) on the
original device. `swap_tensors` swaps the full tensor contents in-place, preserving the parameter's identity so
that any dict keyed by `id(param)` remains valid.
Refer to https://github.com/huggingface/diffusers/pull/13276#discussion_r2944471548 for the full discussion.
"""
torch.utils.swap_tensors(param, source)
def _restore_torchao_tensor(param: torch.Tensor, source: torch.Tensor) -> None:
"""Restore internal tensor data of a TorchAO parameter from `source` without mutating `source`.
Unlike `_swap_torchao_tensor` this copies attribute references one-by-one via `setattr` so that `source` is **not**
modified. Use this when `source` is a cached tensor that must remain unchanged (e.g. a pinned CPU copy in
`cpu_param_dict`).
"""
for attr_name in _get_torchao_inner_tensor_names(source):
setattr(param, attr_name, getattr(source, attr_name))
def _record_stream_torchao_tensor(param: torch.Tensor, stream) -> None:
"""Record stream for all internal tensors of a TorchAO parameter."""
for attr_name in _get_torchao_inner_tensor_names(param):
getattr(param, attr_name).record_stream(stream)
# fmt: off
_GROUP_OFFLOADING = "group_offloading"
_LAYER_EXECUTION_TRACKER = "layer_execution_tracker"
_LAZY_PREFETCH_GROUP_OFFLOADING = "lazy_prefetch_group_offloading"
_GROUP_ID_LAZY_LEAF = "lazy_leafs"
# fmt: on
class GroupOffloadingType(str, Enum):
BLOCK_LEVEL = "block_level"
LEAF_LEVEL = "leaf_level"
@dataclass
class GroupOffloadingConfig:
onload_device: torch.device
offload_device: torch.device
offload_type: GroupOffloadingType
non_blocking: bool
record_stream: bool
low_cpu_mem_usage: bool
num_blocks_per_group: int | None = None
offload_to_disk_path: str | None = None
stream: torch.cuda.Stream | torch.Stream | None = None
block_modules: list[str] | None = None
exclude_kwargs: list[str] | None = None
module_prefix: str = ""
class ModuleGroup:
def __init__(
self,
modules: list[torch.nn.Module],
offload_device: torch.device,
onload_device: torch.device,
offload_leader: torch.nn.Module,
onload_leader: torch.nn.Module | None = None,
parameters: list[torch.nn.Parameter] | None = None,
buffers: list[torch.Tensor] | None = None,
non_blocking: bool = False,
stream: torch.cuda.Stream | torch.Stream | None = None,
record_stream: bool | None = False,
low_cpu_mem_usage: bool = False,
onload_self: bool = True,
offload_to_disk_path: str | None = None,
group_id: int | str | None = None,
) -> None:
self.modules = modules
self.offload_device = offload_device
self.onload_device = onload_device
self.offload_leader = offload_leader
self.onload_leader = onload_leader
self.parameters = parameters or []
self.buffers = buffers or []
self.non_blocking = non_blocking or stream is not None
self.stream = stream
self.record_stream = record_stream
self.onload_self = onload_self
self.low_cpu_mem_usage = low_cpu_mem_usage
self.offload_to_disk_path = offload_to_disk_path
self._is_offloaded_to_disk = False
if self.offload_to_disk_path is not None:
# Instead of `group_id or str(id(self))` we do this because `group_id` can be "" as well.
self.group_id = group_id if group_id is not None else str(id(self))
short_hash = _compute_group_hash(self.group_id)
self.safetensors_file_path = os.path.join(self.offload_to_disk_path, f"group_{short_hash}.safetensors")
all_tensors = []
for module in self.modules:
all_tensors.extend(list(module.parameters()))
all_tensors.extend(list(module.buffers()))
all_tensors.extend(self.parameters)
all_tensors.extend(self.buffers)
all_tensors = list(dict.fromkeys(all_tensors)) # Remove duplicates
self.tensor_to_key = {tensor: f"tensor_{i}" for i, tensor in enumerate(all_tensors)}
self.key_to_tensor = {v: k for k, v in self.tensor_to_key.items()}
self.cpu_param_dict = {}
else:
self.cpu_param_dict = self._init_cpu_param_dict()
self._torch_accelerator_module = (
getattr(torch, torch.accelerator.current_accelerator().type)
if hasattr(torch, "accelerator")
else torch.cuda
)
@staticmethod
def _to_cpu(tensor, low_cpu_mem_usage):
# For TorchAO tensors, `.data` returns an incomplete wrapper without internal attributes
# (e.g. `.qdata`, `.scale`), so we must call `.cpu()` on the tensor directly.
t = tensor.cpu() if _is_torchao_tensor(tensor) else tensor.data.cpu()
return t if low_cpu_mem_usage else t.pin_memory()
def _init_cpu_param_dict(self):
cpu_param_dict = {}
if self.stream is None:
return cpu_param_dict
for module in self.modules:
for param in module.parameters():
cpu_param_dict[param] = self._to_cpu(param, self.low_cpu_mem_usage)
for buffer in module.buffers():
cpu_param_dict[buffer] = self._to_cpu(buffer, self.low_cpu_mem_usage)
for param in self.parameters:
cpu_param_dict[param] = self._to_cpu(param, self.low_cpu_mem_usage)
for buffer in self.buffers:
cpu_param_dict[buffer] = self._to_cpu(buffer, self.low_cpu_mem_usage)
return cpu_param_dict
@contextmanager
def _pinned_memory_tensors(self):
try:
pinned_dict = {
param: tensor.pin_memory() if not tensor.is_pinned() else tensor
for param, tensor in self.cpu_param_dict.items()
}
yield pinned_dict
finally:
pinned_dict = None
def _transfer_tensor_to_device(self, tensor, source_tensor, default_stream):
moved = source_tensor.to(self.onload_device, non_blocking=self.non_blocking)
if _is_torchao_tensor(tensor):
_swap_torchao_tensor(tensor, moved)
else:
tensor.data = moved
if self.record_stream:
if _is_torchao_tensor(tensor):
_record_stream_torchao_tensor(tensor, default_stream)
else:
tensor.data.record_stream(default_stream)
def _process_tensors_from_modules(self, pinned_memory=None, default_stream=None):
for group_module in self.modules:
for param in group_module.parameters():
source = pinned_memory[param] if pinned_memory else param.data
self._transfer_tensor_to_device(param, source, default_stream)
for buffer in group_module.buffers():
source = pinned_memory[buffer] if pinned_memory else buffer.data
self._transfer_tensor_to_device(buffer, source, default_stream)
for param in self.parameters:
source = pinned_memory[param] if pinned_memory else param.data
self._transfer_tensor_to_device(param, source, default_stream)
for buffer in self.buffers:
source = pinned_memory[buffer] if pinned_memory else buffer.data
self._transfer_tensor_to_device(buffer, source, default_stream)
def _check_disk_offload_torchao(self):
all_tensors = list(self.tensor_to_key.keys())
has_torchao = any(_is_torchao_tensor(t) for t in all_tensors)
if has_torchao:
raise ValueError(
"Disk offloading is not supported for TorchAO quantized tensors because safetensors "
"cannot serialize TorchAO subclass tensors. Use memory offloading instead by not "
"setting `offload_to_disk_path`."
)
def _onload_from_disk(self):
self._check_disk_offload_torchao()
if self.stream is not None:
# Wait for previous Host->Device transfer to complete
self.stream.synchronize()
context = nullcontext() if self.stream is None else self._torch_accelerator_module.stream(self.stream)
current_stream = self._torch_accelerator_module.current_stream() if self.record_stream else None
with context:
if self.stream is not None:
# Load to CPU first, pin memory, then async copy to the target device
loaded_tensors = safetensors.torch.load_file(self.safetensors_file_path, device="cpu")
for key, tensor_obj in self.key_to_tensor.items():
pinned_tensor = loaded_tensors[key].pin_memory()
tensor_obj.data = pinned_tensor.to(self.onload_device, non_blocking=self.non_blocking)
if self.record_stream:
tensor_obj.data.record_stream(current_stream)
else:
# Load directly to the target device
onload_device = (
self.onload_device.type if isinstance(self.onload_device, torch.device) else self.onload_device
)
loaded_tensors = safetensors.torch.load_file(self.safetensors_file_path, device=onload_device)
for key, tensor_obj in self.key_to_tensor.items():
tensor_obj.data = loaded_tensors[key]
def _onload_from_memory(self):
if self.stream is not None:
# Wait for previous Host->Device transfer to complete
self.stream.synchronize()
context = nullcontext() if self.stream is None else self._torch_accelerator_module.stream(self.stream)
default_stream = self._torch_accelerator_module.current_stream() if self.stream is not None else None
with context:
if self.stream is not None:
with self._pinned_memory_tensors() as pinned_memory:
self._process_tensors_from_modules(pinned_memory, default_stream=default_stream)
else:
self._process_tensors_from_modules(None)
def _offload_to_disk(self):
self._check_disk_offload_torchao()
# TODO: we can potentially optimize this code path by checking if the _all_ the desired
# safetensor files exist on the disk and if so, skip this step entirely, reducing IO
# overhead. Currently, we just check if the given `safetensors_file_path` exists and if not
# we perform a write.
# Check if the file has been saved in this session or if it already exists on disk.
if not self._is_offloaded_to_disk and not os.path.exists(self.safetensors_file_path):
os.makedirs(os.path.dirname(self.safetensors_file_path), exist_ok=True)
tensors_to_save = {key: tensor.data.to(self.offload_device) for tensor, key in self.tensor_to_key.items()}
safetensors.torch.save_file(tensors_to_save, self.safetensors_file_path)
# The group is now considered offloaded to disk for the rest of the session.
self._is_offloaded_to_disk = True
# We do this to free up the RAM which is still holding the up tensor data.
for tensor_obj in self.tensor_to_key.keys():
tensor_obj.data = torch.empty_like(tensor_obj.data, device=self.offload_device)
def _offload_to_memory(self):
if self.stream is not None:
if not self.record_stream:
self._torch_accelerator_module.current_stream().synchronize()
for group_module in self.modules:
for param in group_module.parameters():
if _is_torchao_tensor(param):
_restore_torchao_tensor(param, self.cpu_param_dict[param])
else:
param.data = self.cpu_param_dict[param]
for param in self.parameters:
if _is_torchao_tensor(param):
_restore_torchao_tensor(param, self.cpu_param_dict[param])
else:
param.data = self.cpu_param_dict[param]
for buffer in self.buffers:
if _is_torchao_tensor(buffer):
_restore_torchao_tensor(buffer, self.cpu_param_dict[buffer])
else:
buffer.data = self.cpu_param_dict[buffer]
else:
for group_module in self.modules:
group_module.to(self.offload_device, non_blocking=False)
for param in self.parameters:
if _is_torchao_tensor(param):
moved = param.to(self.offload_device, non_blocking=False)
_swap_torchao_tensor(param, moved)
else:
param.data = param.data.to(self.offload_device, non_blocking=False)
for buffer in self.buffers:
if _is_torchao_tensor(buffer):
moved = buffer.to(self.offload_device, non_blocking=False)
_swap_torchao_tensor(buffer, moved)
else:
buffer.data = buffer.data.to(self.offload_device, non_blocking=False)
@torch.compiler.disable()
def onload_(self):
r"""Onloads the group of parameters to the onload_device."""
if self.offload_to_disk_path is not None:
self._onload_from_disk()
else:
self._onload_from_memory()
@torch.compiler.disable()
def offload_(self):
r"""Offloads the group of parameters to the offload_device."""
if self.offload_to_disk_path:
self._offload_to_disk()
else:
self._offload_to_memory()
class GroupOffloadingHook(ModelHook):
r"""
A hook that offloads groups of torch.nn.Module to the CPU for storage and onloads to accelerator device for
computation. Each group has one "onload leader" module that is responsible for onloading, and an "offload leader"
module that is responsible for offloading. If prefetching is enabled, the onload leader of the previous module
group is responsible for onloading the current module group.
"""
_is_stateful = False
def __init__(self, group: ModuleGroup, *, config: GroupOffloadingConfig) -> None:
self.group = group
self.next_group: ModuleGroup | None = None
self.config = config
def initialize_hook(self, module: torch.nn.Module) -> torch.nn.Module:
if self.group.offload_leader == module:
self.group.offload_()
return module
def pre_forward(self, module: torch.nn.Module, *args, **kwargs):
# If there wasn't an onload_leader assigned, we assume that the submodule that first called its forward
# method is the onload_leader of the group.
if self.group.onload_leader is None:
self.group.onload_leader = module
# If the current module is the onload_leader of the group, we onload the group if it is supposed
# to onload itself. In the case of using prefetching with streams, we onload the next group if
# it is not supposed to onload itself.
if self.group.onload_leader == module:
if self.group.onload_self:
self.group.onload_()
else:
# onload_self=False means this group relies on prefetching from a previous group.
# However, for conditionally-executed modules (e.g. patch_short/patch_mid/patch_long in Helios),
# the prefetch chain may not cover them if they were absent during the first forward pass
# when the execution order was traced. In that case, their weights remain on offload_device,
# so we fall back to a synchronous onload here.
params = [p for m in self.group.modules for p in m.parameters()] + list(self.group.parameters)
if params and params[0].device == self.group.offload_device:
self.group.onload_()
if self.group.stream is not None:
self.group.stream.synchronize()
should_onload_next_group = self.next_group is not None and not self.next_group.onload_self
if should_onload_next_group:
self.next_group.onload_()
should_synchronize = (
not self.group.onload_self and self.group.stream is not None and not should_onload_next_group
)
if should_synchronize:
# If this group didn't onload itself, it means it was asynchronously onloaded by the
# previous group. We need to synchronize the side stream to ensure parameters
# are completely loaded to proceed with forward pass. Without this, uninitialized
# weights will be used in the computation, leading to incorrect results
# Also, we should only do this synchronization if we don't already do it from the sync call in
# self.next_group.onload_, hence the `not should_onload_next_group` check.
self.group.stream.synchronize()
args = send_to_device(args, self.group.onload_device, non_blocking=self.group.non_blocking)
# Some Autoencoder models use a feature cache that is passed through submodules
# and modified in place. The `send_to_device` call returns a copy of this feature cache object
# which breaks the inplace updates. Use `exclude_kwargs` to mark these cache features
exclude_kwargs = self.config.exclude_kwargs or []
if exclude_kwargs:
moved_kwargs = send_to_device(
{k: v for k, v in kwargs.items() if k not in exclude_kwargs},
self.group.onload_device,
non_blocking=self.group.non_blocking,
)
kwargs.update(moved_kwargs)
else:
kwargs = send_to_device(kwargs, self.group.onload_device, non_blocking=self.group.non_blocking)
return args, kwargs
def post_forward(self, module: torch.nn.Module, output):
if self.group.offload_leader == module:
self.group.offload_()
return output
class LazyPrefetchGroupOffloadingHook(ModelHook):
r"""
A hook, used in conjunction with GroupOffloadingHook, that applies lazy prefetching to groups of torch.nn.Module.
This hook is used to determine the order in which the layers are executed during the forward pass. Once the layer
invocation order is known, assignments of the next_group attribute for prefetching can be made, which allows
prefetching groups in the correct order.
"""
_is_stateful = False
def __init__(self):
self.execution_order: list[tuple[str, torch.nn.Module]] = []
self._layer_execution_tracker_module_names = set()
def initialize_hook(self, module):
def make_execution_order_update_callback(current_name, current_submodule):
def callback():
if not torch.compiler.is_compiling():
logger.debug(f"Adding {current_name} to the execution order")
self.execution_order.append((current_name, current_submodule))
return callback
# To every submodule that contains a group offloading hook (at this point, no prefetching is enabled for any
# of the groups), we add a layer execution tracker hook that will be used to determine the order in which the
# layers are executed during the forward pass.
for name, submodule in module.named_modules():
if name == "" or not hasattr(submodule, "_diffusers_hook"):
continue
registry = HookRegistry.check_if_exists_or_initialize(submodule)
group_offloading_hook = registry.get_hook(_GROUP_OFFLOADING)
if group_offloading_hook is not None:
# For the first forward pass, we have to load in a blocking manner
group_offloading_hook.group.non_blocking = False
layer_tracker_hook = LayerExecutionTrackerHook(make_execution_order_update_callback(name, submodule))
registry.register_hook(layer_tracker_hook, _LAYER_EXECUTION_TRACKER)
self._layer_execution_tracker_module_names.add(name)
return module
def post_forward(self, module, output):
# At this point, for the current modules' submodules, we know the execution order of the layers. We can now
# remove the layer execution tracker hooks and apply prefetching by setting the next_group attribute for each
# group offloading hook.
num_executed = len(self.execution_order)
execution_order_module_names = {name for name, _ in self.execution_order}
# It may be possible that some layers were not executed during the forward pass. This can happen if the layer
# is not used in the forward pass, or if the layer is not executed due to some other reason. In such cases, we
# may not be able to apply prefetching in the correct order, which can lead to device-mismatch related errors
# if the missing layers end up being executed in the future.
if execution_order_module_names != self._layer_execution_tracker_module_names:
unexecuted_layers = list(self._layer_execution_tracker_module_names - execution_order_module_names)
if not torch.compiler.is_compiling():
logger.warning(
"It seems like some layers were not executed during the forward pass. This may lead to problems when "
"applying lazy prefetching with automatic tracing and lead to device-mismatch related errors. Please "
"make sure that all layers are executed during the forward pass. The following layers were not executed:\n"
f"{unexecuted_layers=}"
)
# Remove the layer execution tracker hooks from the submodules
base_module_registry = module._diffusers_hook
registries = [submodule._diffusers_hook for _, submodule in self.execution_order]
group_offloading_hooks = [registry.get_hook(_GROUP_OFFLOADING) for registry in registries]
for i in range(num_executed):
registries[i].remove_hook(_LAYER_EXECUTION_TRACKER, recurse=False)
# Remove the current lazy prefetch group offloading hook so that it doesn't interfere with the next forward pass
base_module_registry.remove_hook(_LAZY_PREFETCH_GROUP_OFFLOADING, recurse=False)
# LazyPrefetchGroupOffloadingHook is only used with streams, so we know that non_blocking should be True.
# We disable non_blocking for the first forward pass, but need to enable it for the subsequent passes to
# see the benefits of prefetching.
for hook in group_offloading_hooks:
hook.group.non_blocking = True
# Set required attributes for prefetching
if num_executed > 0:
base_module_group_offloading_hook = base_module_registry.get_hook(_GROUP_OFFLOADING)
base_module_group_offloading_hook.next_group = group_offloading_hooks[0].group
base_module_group_offloading_hook.next_group.onload_self = False
for i in range(num_executed - 1):
name1, _ = self.execution_order[i]
name2, _ = self.execution_order[i + 1]
if not torch.compiler.is_compiling():
logger.debug(f"Applying lazy prefetch group offloading from {name1} to {name2}")
group_offloading_hooks[i].next_group = group_offloading_hooks[i + 1].group
group_offloading_hooks[i].next_group.onload_self = False
return output
class LayerExecutionTrackerHook(ModelHook):
r"""
A hook that tracks the order in which the layers are executed during the forward pass by calling back to the
LazyPrefetchGroupOffloadingHook to update the execution order.
"""
_is_stateful = False
def __init__(self, execution_order_update_callback):
self.execution_order_update_callback = execution_order_update_callback
def pre_forward(self, module, *args, **kwargs):
self.execution_order_update_callback()
return args, kwargs
def apply_group_offloading(
module: torch.nn.Module,
onload_device: str | torch.device,
offload_device: str | torch.device = torch.device("cpu"),
offload_type: str | GroupOffloadingType = "block_level",
num_blocks_per_group: int | None = None,
non_blocking: bool = False,
use_stream: bool = False,
record_stream: bool = False,
low_cpu_mem_usage: bool = False,
offload_to_disk_path: str | None = None,
block_modules: list[str] | None = None,
exclude_kwargs: list[str] | None = None,
) -> None:
r"""
Applies group offloading to the internal layers of a torch.nn.Module. To understand what group offloading is, and
where it is beneficial, we need to first provide some context on how other supported offloading methods work.
Typically, offloading is done at two levels:
- Module-level: In Diffusers, this can be enabled using the `ModelMixin::enable_model_cpu_offload()` method. It
works by offloading each component of a pipeline to the CPU for storage, and onloading to the accelerator device
when needed for computation. This method is more memory-efficient than keeping all components on the accelerator,
but the memory requirements are still quite high. For this method to work, one needs memory equivalent to size of
the model in runtime dtype + size of largest intermediate activation tensors to be able to complete the forward
pass.
- Leaf-level: In Diffusers, this can be enabled using the `ModelMixin::enable_sequential_cpu_offload()` method. It
works by offloading the lowest leaf-level parameters of the computation graph to the CPU for storage, and
onloading only the leafs to the accelerator device for computation. This uses the lowest amount of accelerator
memory, but can be slower due to the excessive number of device synchronizations.
Group offloading is a middle ground between the two methods. It works by offloading groups of internal layers,
(either `torch.nn.ModuleList` or `torch.nn.Sequential`). This method uses lower memory than module-level
offloading. It is also faster than leaf-level/sequential offloading, as the number of device synchronizations is
reduced.
Another supported feature (for CUDA devices with support for asynchronous data transfer streams) is the ability to
overlap data transfer and computation to reduce the overall execution time compared to sequential offloading. This
is enabled using layer prefetching with streams, i.e., the layer that is to be executed next starts onloading to
the accelerator device while the current layer is being executed - this increases the memory requirements slightly.
Note that this implementation also supports leaf-level offloading but can be made much faster when using streams.
Args:
module (`torch.nn.Module`):
The module to which group offloading is applied.
onload_device (`torch.device`):
The device to which the group of modules are onloaded.
offload_device (`torch.device`, defaults to `torch.device("cpu")`):
The device to which the group of modules are offloaded. This should typically be the CPU. Default is CPU.
offload_type (`str` or `GroupOffloadingType`, defaults to "block_level"):
The type of offloading to be applied. Can be one of "block_level" or "leaf_level". Default is
"block_level".
offload_to_disk_path (`str`, *optional*, defaults to `None`):
The path to the directory where parameters will be offloaded. Setting this option can be useful in limited
RAM environment settings where a reasonable speed-memory trade-off is desired.
num_blocks_per_group (`int`, *optional*):
The number of blocks per group when using offload_type="block_level". This is required when using
offload_type="block_level".
non_blocking (`bool`, defaults to `False`):
If True, offloading and onloading is done with non-blocking data transfer.
use_stream (`bool`, defaults to `False`):
If True, offloading and onloading is done asynchronously using a CUDA stream. This can be useful for
overlapping computation and data transfer.
record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor
as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to the
[PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) more
details.
low_cpu_mem_usage (`bool`, defaults to `False`):
If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. This
option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
the CPU memory is a bottleneck but may counteract the benefits of using streams.
block_modules (`list[str]`, *optional*):
List of module names that should be treated as blocks for offloading. If provided, only these modules will
be considered for block-level offloading. If not provided, the default block detection logic will be used.
exclude_kwargs (`list[str]`, *optional*):
List of kwarg keys that should not be processed by send_to_device. This is useful for mutable state like
caching lists that need to maintain their object identity across forward passes. If not provided, will be
inferred from the module's `_skip_keys` attribute if it exists.
Example:
```python
>>> from diffusers import CogVideoXTransformer3DModel
>>> from diffusers.hooks import apply_group_offloading
>>> transformer = CogVideoXTransformer3DModel.from_pretrained(
... "THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16
... )
>>> apply_group_offloading(
... transformer,
... onload_device=torch.device("cuda"),
... offload_device=torch.device("cpu"),
... offload_type="block_level",
... num_blocks_per_group=2,
... use_stream=True,
... )
```
"""
onload_device = torch.device(onload_device) if isinstance(onload_device, str) else onload_device
offload_device = torch.device(offload_device) if isinstance(offload_device, str) else offload_device
offload_type = GroupOffloadingType(offload_type)
stream = None
if use_stream:
if torch.cuda.is_available():
stream = torch.cuda.Stream()
elif hasattr(torch, "xpu") and torch.xpu.is_available():
stream = torch.Stream()
else:
raise ValueError("Using streams for data transfer requires a CUDA device, or an Intel XPU device.")
if not use_stream and record_stream:
raise ValueError("`record_stream` cannot be True when `use_stream=False`.")
if offload_type == GroupOffloadingType.BLOCK_LEVEL and num_blocks_per_group is None:
raise ValueError("`num_blocks_per_group` must be provided when using `offload_type='block_level'.")
_raise_error_if_accelerate_model_or_sequential_hook_present(module)
if block_modules is None:
block_modules = getattr(module, "_group_offload_block_modules", None)
if exclude_kwargs is None:
exclude_kwargs = getattr(module, "_skip_keys", None)
config = GroupOffloadingConfig(
onload_device=onload_device,
offload_device=offload_device,
offload_type=offload_type,
num_blocks_per_group=num_blocks_per_group,
non_blocking=non_blocking,
stream=stream,
record_stream=record_stream,
low_cpu_mem_usage=low_cpu_mem_usage,
offload_to_disk_path=offload_to_disk_path,
block_modules=block_modules,
exclude_kwargs=exclude_kwargs,
)
_apply_group_offloading(module, config)
def _apply_group_offloading(module: torch.nn.Module, config: GroupOffloadingConfig) -> None:
if config.offload_type == GroupOffloadingType.BLOCK_LEVEL:
_apply_group_offloading_block_level(module, config)
elif config.offload_type == GroupOffloadingType.LEAF_LEVEL:
_apply_group_offloading_leaf_level(module, config)
else:
assert False
def _apply_group_offloading_block_level(module: torch.nn.Module, config: GroupOffloadingConfig) -> None:
r"""
This function applies offloading to groups of torch.nn.ModuleList or torch.nn.Sequential blocks, and explicitly
defined block modules. In comparison to the "leaf_level" offloading, which is more fine-grained, this offloading is
done at the top-level blocks and modules specified in block_modules.
When block_modules is provided, only those modules will be treated as blocks for offloading. For each specified
module, recursively apply block offloading to it.
"""
if config.stream is not None and config.num_blocks_per_group != 1:
logger.warning(
f"Using streams is only supported for num_blocks_per_group=1. Got {config.num_blocks_per_group=}. Setting it to 1."
)
config.num_blocks_per_group = 1
block_modules = set(config.block_modules) if config.block_modules is not None else set()
# Create module groups for ModuleList and Sequential blocks, and explicitly defined block modules
modules_with_group_offloading = set()
unmatched_modules = []
matched_module_groups = []
for name, submodule in module.named_children():
# Check if this is an explicitly defined block module
if name in block_modules:
# Track submodule using a prefix to avoid filename collisions during disk offload.
# Without this, submodules sharing the same model class would be assigned identical
# filenames (derived from the class name).
prefix = f"{config.module_prefix}{name}." if config.module_prefix else f"{name}."
submodule_config = replace(config, module_prefix=prefix)
_apply_group_offloading_block_level(submodule, submodule_config)
modules_with_group_offloading.add(name)
elif isinstance(submodule, (torch.nn.ModuleList, torch.nn.Sequential)):
# Handle ModuleList and Sequential blocks as before
for i in range(0, len(submodule), config.num_blocks_per_group):
current_modules = list(submodule[i : i + config.num_blocks_per_group])
if len(current_modules) == 0:
continue
group_id = f"{config.module_prefix}{name}_{i}_{i + len(current_modules) - 1}"
group = ModuleGroup(
modules=current_modules,
offload_device=config.offload_device,
onload_device=config.onload_device,
offload_to_disk_path=config.offload_to_disk_path,
offload_leader=current_modules[-1],
onload_leader=current_modules[0],
non_blocking=config.non_blocking,
stream=config.stream,
record_stream=config.record_stream,
low_cpu_mem_usage=config.low_cpu_mem_usage,
onload_self=True,
group_id=group_id,
)
matched_module_groups.append(group)
for j in range(i, i + len(current_modules)):
modules_with_group_offloading.add(f"{name}.{j}")
else:
# This is an unmatched module
unmatched_modules.append((name, submodule))
# Apply group offloading hooks to the module groups
for i, group in enumerate(matched_module_groups):
for group_module in group.modules:
_apply_group_offloading_hook(group_module, group, config=config)
# Parameters and Buffers of the top-level module need to be offloaded/onloaded separately
# when the forward pass of this module is called. This is because the top-level module is not
# part of any group (as doing so would lead to no VRAM savings).
parameters = _gather_parameters_with_no_group_offloading_parent(module, modules_with_group_offloading)
buffers = _gather_buffers_with_no_group_offloading_parent(module, modules_with_group_offloading)
parameters = [param for _, param in parameters]
buffers = [buffer for _, buffer in buffers]
# Create a group for the remaining unmatched submodules of the top-level
# module so that they are on the correct device when the forward pass is called.
unmatched_modules = [unmatched_module for _, unmatched_module in unmatched_modules]
if len(unmatched_modules) > 0 or len(parameters) > 0 or len(buffers) > 0:
unmatched_group = ModuleGroup(
modules=unmatched_modules,
offload_device=config.offload_device,
onload_device=config.onload_device,
offload_to_disk_path=config.offload_to_disk_path,
offload_leader=module,
onload_leader=module,
parameters=parameters,
buffers=buffers,
non_blocking=False,
stream=None,
record_stream=False,
onload_self=True,
group_id=f"{config.module_prefix}{module.__class__.__name__}_unmatched_group",
)
if config.stream is None:
_apply_group_offloading_hook(module, unmatched_group, config=config)
else:
_apply_lazy_group_offloading_hook(module, unmatched_group, config=config)
def _apply_group_offloading_leaf_level(module: torch.nn.Module, config: GroupOffloadingConfig) -> None:
r"""
This function applies offloading to groups of leaf modules in a torch.nn.Module. This method has minimal memory
requirements. However, it can be slower compared to other offloading methods due to the excessive number of device
synchronizations. When using devices that support streams to overlap data transfer and computation, this method can
reduce memory usage without any performance degradation.
"""
# Create module groups for leaf modules and apply group offloading hooks
modules_with_group_offloading = set()
for name, submodule in module.named_modules():
if not isinstance(submodule, _GO_LC_SUPPORTED_PYTORCH_LAYERS):
continue
group = ModuleGroup(
modules=[submodule],
offload_device=config.offload_device,
onload_device=config.onload_device,
offload_to_disk_path=config.offload_to_disk_path,
offload_leader=submodule,
onload_leader=submodule,
non_blocking=config.non_blocking,
stream=config.stream,
record_stream=config.record_stream,
low_cpu_mem_usage=config.low_cpu_mem_usage,
onload_self=True,
group_id=name,
)
_apply_group_offloading_hook(submodule, group, config=config)
modules_with_group_offloading.add(name)
# Parameters and Buffers at all non-leaf levels need to be offloaded/onloaded separately when the forward pass
# of the module is called
module_dict = dict(module.named_modules())
parameters = _gather_parameters_with_no_group_offloading_parent(module, modules_with_group_offloading)
buffers = _gather_buffers_with_no_group_offloading_parent(module, modules_with_group_offloading)
# Find closest module parent for each parameter and buffer, and attach group hooks
parent_to_parameters = {}
for name, param in parameters:
parent_name = _find_parent_module_in_module_dict(name, module_dict)
if parent_name in parent_to_parameters:
parent_to_parameters[parent_name].append(param)
else:
parent_to_parameters[parent_name] = [param]
parent_to_buffers = {}
for name, buffer in buffers:
parent_name = _find_parent_module_in_module_dict(name, module_dict)
if parent_name in parent_to_buffers:
parent_to_buffers[parent_name].append(buffer)
else:
parent_to_buffers[parent_name] = [buffer]
parent_names = set(parent_to_parameters.keys()) | set(parent_to_buffers.keys())
for name in parent_names:
parameters = parent_to_parameters.get(name, [])
buffers = parent_to_buffers.get(name, [])
parent_module = module_dict[name]
group = ModuleGroup(
modules=[],
offload_device=config.offload_device,
onload_device=config.onload_device,
offload_leader=parent_module,
onload_leader=parent_module,
offload_to_disk_path=config.offload_to_disk_path,
parameters=parameters,
buffers=buffers,
non_blocking=config.non_blocking,
stream=config.stream,
record_stream=config.record_stream,
low_cpu_mem_usage=config.low_cpu_mem_usage,
onload_self=True,
group_id=name,
)
_apply_group_offloading_hook(parent_module, group, config=config)
if config.stream is not None:
# When using streams, we need to know the layer execution order for applying prefetching (to overlap data transfer
# and computation). Since we don't know the order beforehand, we apply a lazy prefetching hook that will find the
# execution order and apply prefetching in the correct order.
unmatched_group = ModuleGroup(
modules=[],
offload_device=config.offload_device,
onload_device=config.onload_device,
offload_to_disk_path=config.offload_to_disk_path,
offload_leader=module,
onload_leader=module,
parameters=None,
buffers=None,
non_blocking=False,
stream=None,
record_stream=False,
low_cpu_mem_usage=config.low_cpu_mem_usage,
onload_self=True,
group_id=_GROUP_ID_LAZY_LEAF,
)
_apply_lazy_group_offloading_hook(module, unmatched_group, config=config)
def _apply_group_offloading_hook(
module: torch.nn.Module,
group: ModuleGroup,
*,
config: GroupOffloadingConfig,
) -> None:
registry = HookRegistry.check_if_exists_or_initialize(module)
# We may have already registered a group offloading hook if the module had a torch.nn.Parameter whose parent
# is the current module. In such cases, we don't want to overwrite the existing group offloading hook.
if registry.get_hook(_GROUP_OFFLOADING) is None:
hook = GroupOffloadingHook(group, config=config)
registry.register_hook(hook, _GROUP_OFFLOADING)
def _apply_lazy_group_offloading_hook(
module: torch.nn.Module,
group: ModuleGroup,
*,
config: GroupOffloadingConfig,
) -> None:
registry = HookRegistry.check_if_exists_or_initialize(module)
# We may have already registered a group offloading hook if the module had a torch.nn.Parameter whose parent
# is the current module. In such cases, we don't want to overwrite the existing group offloading hook.
if registry.get_hook(_GROUP_OFFLOADING) is None:
hook = GroupOffloadingHook(group, config=config)
registry.register_hook(hook, _GROUP_OFFLOADING)
lazy_prefetch_hook = LazyPrefetchGroupOffloadingHook()
registry.register_hook(lazy_prefetch_hook, _LAZY_PREFETCH_GROUP_OFFLOADING)
def _gather_parameters_with_no_group_offloading_parent(
module: torch.nn.Module, modules_with_group_offloading: Set[str]
) -> list[torch.nn.Parameter]:
parameters = []
for name, parameter in module.named_parameters():
has_parent_with_group_offloading = False
atoms = name.split(".")
while len(atoms) > 0:
parent_name = ".".join(atoms)
if parent_name in modules_with_group_offloading:
has_parent_with_group_offloading = True
break
atoms.pop()
if not has_parent_with_group_offloading:
parameters.append((name, parameter))
return parameters
def _gather_buffers_with_no_group_offloading_parent(
module: torch.nn.Module, modules_with_group_offloading: Set[str]
) -> list[torch.Tensor]:
buffers = []
for name, buffer in module.named_buffers():
has_parent_with_group_offloading = False
atoms = name.split(".")
while len(atoms) > 0:
parent_name = ".".join(atoms)
if parent_name in modules_with_group_offloading:
has_parent_with_group_offloading = True
break
atoms.pop()
if not has_parent_with_group_offloading:
buffers.append((name, buffer))
return buffers
def _find_parent_module_in_module_dict(name: str, module_dict: dict[str, torch.nn.Module]) -> str:
atoms = name.split(".")
while len(atoms) > 0:
parent_name = ".".join(atoms)
if parent_name in module_dict:
return parent_name
atoms.pop()
return ""
def _raise_error_if_accelerate_model_or_sequential_hook_present(module: torch.nn.Module) -> None:
if not is_accelerate_available():
return
for name, submodule in module.named_modules():
if not hasattr(submodule, "_hf_hook"):
continue
if isinstance(submodule._hf_hook, (AlignDevicesHook, CpuOffload)):
raise ValueError(
f"Cannot apply group offloading to a module that is already applying an alternative "