-
Notifications
You must be signed in to change notification settings - Fork 12.6k
Expand file tree
/
Copy pathmodel_management.py
More file actions
1828 lines (1502 loc) · 57.1 KB
/
model_management.py
File metadata and controls
1828 lines (1502 loc) · 57.1 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
"""
This file is part of ComfyUI.
Copyright (C) 2024 Comfy
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
"""
import psutil
import logging
from enum import Enum
from comfy.cli_args import args, PerformanceFeature
import threading
import torch
import sys
import platform
import weakref
import gc
import os
from contextlib import nullcontext
import comfy.memory_management
import comfy.utils
import comfy.quant_ops
class VRAMState(Enum):
DISABLED = 0 #No vram present: no need to move models to vram
NO_VRAM = 1 #Very low vram: enable all the options to save vram
LOW_VRAM = 2
NORMAL_VRAM = 3
HIGH_VRAM = 4
SHARED = 5 #No dedicated vram: memory shared between CPU and GPU but models still need to be moved between both.
class CPUState(Enum):
GPU = 0
CPU = 1
MPS = 2
# Determine VRAM State
vram_state = VRAMState.NORMAL_VRAM
set_vram_to = VRAMState.NORMAL_VRAM
cpu_state = CPUState.GPU
total_vram = 0
# Training Related State
in_training = False
training_fp8_bwd = False
def get_supported_float8_types():
float8_types = []
try:
float8_types.append(torch.float8_e4m3fn)
except:
pass
try:
float8_types.append(torch.float8_e4m3fnuz)
except:
pass
try:
float8_types.append(torch.float8_e5m2)
except:
pass
try:
float8_types.append(torch.float8_e5m2fnuz)
except:
pass
try:
float8_types.append(torch.float8_e8m0fnu)
except:
pass
return float8_types
FLOAT8_TYPES = get_supported_float8_types()
xpu_available = False
torch_version = ""
try:
torch_version = torch.version.__version__
temp = torch_version.split(".")
torch_version_numeric = (int(temp[0]), int(temp[1]))
except:
pass
lowvram_available = True
if args.deterministic:
logging.info("Using deterministic algorithms for pytorch")
torch.use_deterministic_algorithms(True, warn_only=True)
directml_enabled = False
if args.directml is not None:
logging.warning("WARNING: torch-directml barely works, is very slow, has not been updated in over 1 year and might be removed soon, please don't use it, there are better options.")
import torch_directml
directml_enabled = True
device_index = args.directml
if device_index < 0:
directml_device = torch_directml.device()
else:
directml_device = torch_directml.device(device_index)
logging.info("Using directml with device: {}".format(torch_directml.device_name(device_index)))
# torch_directml.disable_tiled_resources(True)
lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
try:
import intel_extension_for_pytorch as ipex # noqa: F401
except:
pass
try:
_ = torch.xpu.device_count()
xpu_available = torch.xpu.is_available()
except:
xpu_available = False
try:
if torch.backends.mps.is_available():
cpu_state = CPUState.MPS
import torch.mps
except:
pass
try:
import torch_npu # noqa: F401
_ = torch.npu.device_count()
npu_available = torch.npu.is_available()
except:
npu_available = False
try:
import torch_mlu # noqa: F401
_ = torch.mlu.device_count()
mlu_available = torch.mlu.is_available()
except:
mlu_available = False
try:
ixuca_available = hasattr(torch, "corex")
except:
ixuca_available = False
if args.cpu:
cpu_state = CPUState.CPU
def is_intel_xpu():
global cpu_state
global xpu_available
if cpu_state == CPUState.GPU:
if xpu_available:
return True
return False
def is_ascend_npu():
global npu_available
if npu_available:
return True
return False
def is_mlu():
global mlu_available
if mlu_available:
return True
return False
def is_ixuca():
global ixuca_available
if ixuca_available:
return True
return False
def is_wsl():
version = platform.uname().release
if version.endswith("-Microsoft"):
return True
elif version.endswith("microsoft-standard-WSL2"):
return True
return False
def get_torch_device():
global directml_enabled
global cpu_state
if directml_enabled:
global directml_device
return directml_device
if cpu_state == CPUState.MPS:
return torch.device("mps")
if cpu_state == CPUState.CPU:
return torch.device("cpu")
else:
if is_intel_xpu():
return torch.device("xpu", torch.xpu.current_device())
elif is_ascend_npu():
return torch.device("npu", torch.npu.current_device())
elif is_mlu():
return torch.device("mlu", torch.mlu.current_device())
else:
return torch.device(torch.cuda.current_device())
def get_total_memory(dev=None, torch_total_too=False):
global directml_enabled
if dev is None:
dev = get_torch_device()
if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
mem_total = psutil.virtual_memory().total
mem_total_torch = mem_total
else:
if directml_enabled:
mem_total = 1024 * 1024 * 1024 #TODO
mem_total_torch = mem_total
elif is_intel_xpu():
stats = torch.xpu.memory_stats(dev)
mem_reserved = stats['reserved_bytes.all.current']
mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory
mem_total_torch = mem_reserved
mem_total = mem_total_xpu
elif is_ascend_npu():
stats = torch.npu.memory_stats(dev)
mem_reserved = stats['reserved_bytes.all.current']
_, mem_total_npu = torch.npu.mem_get_info(dev)
mem_total_torch = mem_reserved
mem_total = mem_total_npu
elif is_mlu():
stats = torch.mlu.memory_stats(dev)
mem_reserved = stats['reserved_bytes.all.current']
_, mem_total_mlu = torch.mlu.mem_get_info(dev)
mem_total_torch = mem_reserved
mem_total = mem_total_mlu
else:
stats = torch.cuda.memory_stats(dev)
mem_reserved = stats['reserved_bytes.all.current']
_, mem_total_cuda = torch.cuda.mem_get_info(dev)
mem_total_torch = mem_reserved
mem_total = mem_total_cuda
if torch_total_too:
return (mem_total, mem_total_torch)
else:
return mem_total
def mac_version():
try:
return tuple(int(n) for n in platform.mac_ver()[0].split("."))
except:
return None
total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
total_ram = psutil.virtual_memory().total / (1024 * 1024)
logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
try:
logging.info("pytorch version: {}".format(torch_version))
mac_ver = mac_version()
if mac_ver is not None:
logging.info("Mac Version {}".format(mac_ver))
except:
pass
try:
OOM_EXCEPTION = torch.cuda.OutOfMemoryError
except:
OOM_EXCEPTION = Exception
try:
ACCELERATOR_ERROR = torch.AcceleratorError
except AttributeError:
ACCELERATOR_ERROR = RuntimeError
def is_oom(e):
if isinstance(e, OOM_EXCEPTION):
return True
if isinstance(e, ACCELERATOR_ERROR) and (getattr(e, 'error_code', None) == 2 or "out of memory" in str(e).lower()):
discard_cuda_async_error()
return True
return False
def raise_non_oom(e):
if not is_oom(e):
raise e
XFORMERS_VERSION = ""
XFORMERS_ENABLED_VAE = True
if args.disable_xformers:
XFORMERS_IS_AVAILABLE = False
else:
try:
import xformers
import xformers.ops
XFORMERS_IS_AVAILABLE = True
try:
XFORMERS_IS_AVAILABLE = xformers._has_cpp_library
except:
pass
try:
XFORMERS_VERSION = xformers.version.__version__
logging.info("xformers version: {}".format(XFORMERS_VERSION))
if XFORMERS_VERSION.startswith("0.0.18"):
logging.warning("\nWARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.")
logging.warning("Please downgrade or upgrade xformers to a different version.\n")
XFORMERS_ENABLED_VAE = False
except:
pass
except:
XFORMERS_IS_AVAILABLE = False
def is_nvidia():
global cpu_state
if cpu_state == CPUState.GPU:
if torch.version.cuda:
return True
return False
def is_amd():
global cpu_state
if cpu_state == CPUState.GPU:
if torch.version.hip:
return True
return False
def amd_min_version(device=None, min_rdna_version=0):
if not is_amd():
return False
if is_device_cpu(device):
return False
arch = torch.cuda.get_device_properties(device).gcnArchName
if arch.startswith('gfx') and len(arch) == 7:
try:
cmp_rdna_version = int(arch[4]) + 2
except:
cmp_rdna_version = 0
if cmp_rdna_version >= min_rdna_version:
return True
return False
MIN_WEIGHT_MEMORY_RATIO = 0.4
if is_nvidia():
MIN_WEIGHT_MEMORY_RATIO = 0.0
ENABLE_PYTORCH_ATTENTION = False
if args.use_pytorch_cross_attention:
ENABLE_PYTORCH_ATTENTION = True
XFORMERS_IS_AVAILABLE = False
try:
if is_nvidia():
if torch_version_numeric[0] >= 2:
if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
ENABLE_PYTORCH_ATTENTION = True
if is_intel_xpu() or is_ascend_npu() or is_mlu() or is_ixuca():
if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
ENABLE_PYTORCH_ATTENTION = True
except:
pass
SUPPORT_FP8_OPS = args.supports_fp8_compute
AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
try:
if is_amd():
arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName.split(':')[0]
if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD
logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
try:
rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))
except:
rocm_version = (6, -1)
def aotriton_supported(gpu_arch):
path = torch.__path__[0]
path = os.path.join(os.path.join(path, "lib"), "aotriton.images")
gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path))))
if gpu_arch in gfx:
return True
if "{}x".format(gpu_arch[:-1]) in gfx:
return True
if "{}xx".format(gpu_arch[:-2]) in gfx:
return True
return False
logging.info("AMD arch: {}".format(arch))
logging.info("ROCm version: {}".format(rocm_version))
if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
if aotriton_supported(arch): # AMD efficient attention implementation depends on aotriton.
if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much
if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1150", "gfx1151"]): # TODO: more arches, TODO: gfx950
ENABLE_PYTORCH_ATTENTION = True
if rocm_version >= (7, 0):
if any((a in arch) for a in ["gfx1200", "gfx1201"]):
ENABLE_PYTORCH_ATTENTION = True
if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0
SUPPORT_FP8_OPS = True
except:
pass
if ENABLE_PYTORCH_ATTENTION:
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
PRIORITIZE_FP16 = False # TODO: remove and replace with something that shows exactly which dtype is faster than the other
try:
if (is_nvidia() or is_amd()) and PerformanceFeature.Fp16Accumulation in args.fast:
torch.backends.cuda.matmul.allow_fp16_accumulation = True
PRIORITIZE_FP16 = True # TODO: limit to cards where it actually boosts performance
logging.info("Enabled fp16 accumulation.")
except:
pass
if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
torch.backends.cudnn.benchmark = True
try:
if torch_version_numeric >= (2, 5):
torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
except:
logging.warning("Warning, could not set allow_fp16_bf16_reduction_math_sdp")
if args.lowvram:
set_vram_to = VRAMState.LOW_VRAM
lowvram_available = True
elif args.novram:
set_vram_to = VRAMState.NO_VRAM
elif args.highvram or args.gpu_only:
vram_state = VRAMState.HIGH_VRAM
FORCE_FP32 = False
if args.force_fp32:
logging.info("Forcing FP32, if this improves things please report it.")
FORCE_FP32 = True
if lowvram_available:
if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
vram_state = set_vram_to
if cpu_state != CPUState.GPU:
vram_state = VRAMState.DISABLED
if cpu_state == CPUState.MPS:
vram_state = VRAMState.SHARED
logging.info(f"Set vram state to: {vram_state.name}")
DISABLE_SMART_MEMORY = args.disable_smart_memory
if DISABLE_SMART_MEMORY:
logging.info("Disabling smart memory management")
def get_torch_device_name(device):
if hasattr(device, 'type'):
if device.type == "cuda":
try:
allocator_backend = torch.cuda.get_allocator_backend()
except:
allocator_backend = ""
return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
elif device.type == "xpu":
return "{} {}".format(device, torch.xpu.get_device_name(device))
else:
return "{}".format(device.type)
elif is_intel_xpu():
return "{} {}".format(device, torch.xpu.get_device_name(device))
elif is_ascend_npu():
return "{} {}".format(device, torch.npu.get_device_name(device))
elif is_mlu():
return "{} {}".format(device, torch.mlu.get_device_name(device))
else:
return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))
try:
logging.info("Device: {}".format(get_torch_device_name(get_torch_device())))
except:
logging.warning("Could not pick default device.")
current_loaded_models = []
def module_size(module):
module_mem = 0
sd = module.state_dict()
for k in sd:
t = sd[k]
module_mem += t.nbytes
return module_mem
def module_mmap_residency(module, free=False):
mmap_touched_mem = 0
module_mem = 0
bounced_mmaps = set()
sd = module.state_dict()
for k in sd:
t = sd[k]
module_mem += t.nbytes
storage = t._qdata.untyped_storage() if isinstance(t, comfy.quant_ops.QuantizedTensor) else t.untyped_storage()
if not getattr(storage, "_comfy_tensor_mmap_touched", False):
continue
mmap_touched_mem += t.nbytes
if not free:
continue
storage._comfy_tensor_mmap_touched = False
mmap_obj = storage._comfy_tensor_mmap_refs[0]
if mmap_obj in bounced_mmaps:
continue
mmap_obj.bounce()
bounced_mmaps.add(mmap_obj)
return mmap_touched_mem, module_mem
class LoadedModel:
def __init__(self, model):
self._set_model(model)
self.device = model.load_device
self.real_model = None
self.currently_used = True
self.model_finalizer = None
self._patcher_finalizer = None
def _set_model(self, model):
self._model = weakref.ref(model)
if model.parent is not None:
self._parent_model = weakref.ref(model.parent)
self._patcher_finalizer = weakref.finalize(model, self._switch_parent)
self._patcher_finalizer.atexit = False
def _switch_parent(self):
model = self._parent_model()
if model is not None:
self._set_model(model)
@property
def model(self):
return self._model()
def model_memory(self):
return self.model.model_size()
def model_mmap_residency(self, free=False):
return self.model.model_mmap_residency(free=free)
def model_loaded_memory(self):
return self.model.loaded_size()
def model_offloaded_memory(self):
return self.model.model_size() - self.model.loaded_size()
def model_memory_required(self, device):
if device == self.model.current_loaded_device():
return self.model_offloaded_memory()
else:
return self.model_memory()
def model_load(self, lowvram_model_memory=0, force_patch_weights=False):
self.model.model_patches_to(self.device)
self.model.model_patches_to(self.model.model_dtype())
# if self.model.loaded_size() > 0:
use_more_vram = lowvram_model_memory
if use_more_vram == 0:
use_more_vram = 1e32
self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
real_model = self.model.model
if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
with torch.no_grad():
real_model = ipex.optimize(real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
self.real_model = weakref.ref(real_model)
self.model_finalizer = weakref.finalize(real_model, cleanup_models)
self.model_finalizer.atexit = False
return real_model
def should_reload_model(self, force_patch_weights=False):
if force_patch_weights and self.model.lowvram_patch_counter() > 0:
return True
return False
def model_unload(self, memory_to_free=None, unpatch_weights=True):
if memory_to_free is not None:
if memory_to_free < self.model.loaded_size():
freed = self.model.partially_unload(self.model.offload_device, memory_to_free)
if freed >= memory_to_free:
return False
self.model.detach(unpatch_weights)
self.model_finalizer.detach()
self.model_finalizer = None
self.real_model = None
return True
def model_use_more_vram(self, extra_memory, force_patch_weights=False):
return self.model.partially_load(self.device, extra_memory, force_patch_weights=force_patch_weights)
def __eq__(self, other):
return self.model is other.model
def __del__(self):
if self._patcher_finalizer is not None:
self._patcher_finalizer.detach()
def is_dead(self):
return self.real_model() is not None and self.model is None
def use_more_memory(extra_memory, loaded_models, device):
for m in loaded_models:
if m.device == device:
extra_memory -= m.model_use_more_vram(extra_memory)
if extra_memory <= 0:
break
def offloaded_memory(loaded_models, device):
offloaded_mem = 0
for m in loaded_models:
if m.device == device:
offloaded_mem += m.model_offloaded_memory()
return offloaded_mem
WINDOWS = any(platform.win32_ver())
EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
if WINDOWS:
import comfy.windows
EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
if total_vram > (15 * 1024): # more extra reserved vram on 16GB+ cards
EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
def get_free_ram():
return comfy.windows.get_free_ram()
else:
def get_free_ram():
return psutil.virtual_memory().available
if args.reserve_vram is not None:
EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
logging.debug("Reserving {}MB vram for other applications.".format(EXTRA_RESERVED_VRAM / (1024 * 1024)))
def extra_reserved_memory():
return EXTRA_RESERVED_VRAM
def minimum_inference_memory():
return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
cleanup_models_gc()
unloaded_model = []
can_unload = []
unloaded_models = []
for i in range(len(current_loaded_models) -1, -1, -1):
shift_model = current_loaded_models[i]
if device is None or shift_model.device == device:
if shift_model not in keep_loaded and not shift_model.is_dead():
can_unload.append((-shift_model.model_offloaded_memory(), sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
shift_model.currently_used = False
can_unload_sorted = sorted(can_unload)
for x in can_unload_sorted:
i = x[-1]
memory_to_free = 1e32
pins_to_free = 1e32
if not DISABLE_SMART_MEMORY or device is None:
memory_to_free = 0 if device is None else memory_required - get_free_memory(device)
pins_to_free = pins_required - get_free_ram()
if current_loaded_models[i].model.is_dynamic() and for_dynamic:
#don't actually unload dynamic models for the sake of other dynamic models
#as that works on-demand.
memory_required -= current_loaded_models[i].model.loaded_size()
memory_to_free = 0
if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
unloaded_model.append(i)
if pins_to_free > 0:
logging.debug(f"PIN Unloading {current_loaded_models[i].model.model.__class__.__name__}")
current_loaded_models[i].model.partially_unload_ram(pins_to_free)
for x in can_unload_sorted:
i = x[-1]
ram_to_free = ram_required - psutil.virtual_memory().available
if ram_to_free <= 0 and i not in unloaded_model:
continue
resident_memory, _ = current_loaded_models[i].model_mmap_residency(free=True)
if resident_memory > 0:
logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
for i in sorted(unloaded_model, reverse=True):
unloaded_models.append(current_loaded_models.pop(i))
if len(unloaded_model) > 0:
soft_empty_cache()
elif device is not None:
if vram_state != VRAMState.HIGH_VRAM:
mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
if mem_free_torch > mem_free_total * 0.25:
soft_empty_cache()
return unloaded_models
def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
cleanup_models_gc()
global vram_state
inference_memory = minimum_inference_memory()
extra_mem = max(inference_memory, memory_required + extra_reserved_memory())
if minimum_memory_required is None:
minimum_memory_required = extra_mem
else:
minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
models_temp = set()
for m in models:
models_temp.add(m)
for mm in m.model_patches_models():
models_temp.add(mm)
models = models_temp
models_to_load = []
free_for_dynamic=True
for x in models:
if not x.is_dynamic():
free_for_dynamic = False
loaded_model = LoadedModel(x)
try:
loaded_model_index = current_loaded_models.index(loaded_model)
except:
loaded_model_index = None
if loaded_model_index is not None:
loaded = current_loaded_models[loaded_model_index]
loaded.currently_used = True
models_to_load.append(loaded)
else:
if hasattr(x, "model"):
logging.info(f"Requested to load {x.model.__class__.__name__}")
models_to_load.append(loaded_model)
for loaded_model in models_to_load:
to_unload = []
for i in range(len(current_loaded_models)):
if loaded_model.model.is_clone(current_loaded_models[i].model):
to_unload = [i] + to_unload
for i in to_unload:
model_to_unload = current_loaded_models.pop(i)
model_to_unload.model.detach(unpatch_all=False)
model_to_unload.model_finalizer.detach()
total_memory_required = {}
total_pins_required = {}
total_ram_required = {}
for loaded_model in models_to_load:
device = loaded_model.device
total_memory_required[device] = total_memory_required.get(device, 0) + loaded_model.model_memory_required(device)
resident_memory, model_memory = loaded_model.model_mmap_residency()
pinned_memory = loaded_model.model.pinned_memory_size()
#FIXME: This can over-free the pins as it budgets to pin the entire model. We should
#make this JIT to keep as much pinned as possible.
pins_required = model_memory - pinned_memory
ram_required = model_memory - resident_memory
total_pins_required[device] = total_pins_required.get(device, 0) + pins_required
total_ram_required[device] = total_ram_required.get(device, 0) + ram_required
for device in total_memory_required:
if device != torch.device("cpu"):
free_memory(total_memory_required[device] * 1.1 + extra_mem,
device,
for_dynamic=free_for_dynamic,
pins_required=total_pins_required[device],
ram_required=total_ram_required[device])
for device in total_memory_required:
if device != torch.device("cpu"):
free_mem = get_free_memory(device)
if free_mem < minimum_memory_required:
models_l = free_memory(minimum_memory_required, device, for_dynamic=free_for_dynamic)
logging.info("{} models unloaded.".format(len(models_l)))
for loaded_model in models_to_load:
model = loaded_model.model
torch_dev = model.load_device
if is_device_cpu(torch_dev):
vram_set_state = VRAMState.DISABLED
else:
vram_set_state = vram_state
lowvram_model_memory = 0
if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
loaded_memory = loaded_model.model_loaded_memory()
current_free_mem = get_free_memory(torch_dev) + loaded_memory
lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
lowvram_model_memory = lowvram_model_memory - loaded_memory
if lowvram_model_memory == 0:
lowvram_model_memory = 0.1
if vram_set_state == VRAMState.NO_VRAM:
lowvram_model_memory = 0.1
loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights)
current_loaded_models.insert(0, loaded_model)
return
def load_model_gpu(model):
return load_models_gpu([model])
def loaded_models(only_currently_used=False):
output = []
for m in current_loaded_models:
if only_currently_used:
if not m.currently_used:
continue
output.append(m.model)
return output
def cleanup_models_gc():
do_gc = False
reset_cast_buffers()
for i in range(len(current_loaded_models)):
cur = current_loaded_models[i]
if cur.is_dead():
logging.info("Potential memory leak detected with model {}, doing a full garbage collect, for maximum performance avoid circular references in the model code.".format(cur.real_model().__class__.__name__))
do_gc = True
break
if do_gc:
gc.collect()
soft_empty_cache()
for i in range(len(current_loaded_models)):
cur = current_loaded_models[i]
if cur.is_dead():
logging.warning("WARNING, memory leak with model {}. Please make sure it is not being referenced from somewhere.".format(cur.real_model().__class__.__name__))
def archive_model_dtypes(model):
for name, module in model.named_modules():
for param_name, param in module.named_parameters(recurse=False):
setattr(module, f"{param_name}_comfy_model_dtype", param.dtype)
for buf_name, buf in module.named_buffers(recurse=False):
setattr(module, f"{buf_name}_comfy_model_dtype", buf.dtype)
def cleanup_models():
to_delete = []
for i in range(len(current_loaded_models)):
if current_loaded_models[i].real_model() is None:
to_delete = [i] + to_delete
for i in to_delete:
x = current_loaded_models.pop(i)
del x
def dtype_size(dtype):
dtype_size = 4
if dtype == torch.float16 or dtype == torch.bfloat16:
dtype_size = 2
elif dtype == torch.float32:
dtype_size = 4
else:
try:
dtype_size = dtype.itemsize
except: #Old pytorch doesn't have .itemsize
pass
return dtype_size
def unet_offload_device():
if vram_state == VRAMState.HIGH_VRAM:
return get_torch_device()
else:
return torch.device("cpu")
def unet_inital_load_device(parameters, dtype):
cpu_dev = torch.device("cpu")
if comfy.memory_management.aimdo_enabled:
return cpu_dev
torch_dev = get_torch_device()
if vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.SHARED:
return torch_dev
if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM:
return cpu_dev
model_size = dtype_size(dtype) * parameters
mem_dev = get_free_memory(torch_dev)
mem_cpu = get_free_memory(cpu_dev)
if mem_dev > mem_cpu and model_size < mem_dev:
return torch_dev
else:
return cpu_dev
def maximum_vram_for_weights(device=None):
return (get_total_memory(device) * 0.88 - minimum_inference_memory())
def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32], weight_dtype=None):
if model_params < 0:
model_params = 1000000000000000000000
if args.fp32_unet:
return torch.float32
if args.fp64_unet:
return torch.float64
if args.bf16_unet:
return torch.bfloat16
if args.fp16_unet:
return torch.float16
if args.fp8_e4m3fn_unet:
return torch.float8_e4m3fn
if args.fp8_e5m2_unet:
return torch.float8_e5m2
if args.fp8_e8m0fnu_unet:
return torch.float8_e8m0fnu
fp8_dtype = None
if weight_dtype in FLOAT8_TYPES:
fp8_dtype = weight_dtype
if fp8_dtype is not None:
if supports_fp8_compute(device): #if fp8 compute is supported the casting is most likely not expensive
return fp8_dtype
free_model_memory = maximum_vram_for_weights(device)
if model_params * 2 > free_model_memory:
return fp8_dtype
if PRIORITIZE_FP16 or weight_dtype == torch.float16:
if torch.float16 in supported_dtypes and should_use_fp16(device=device, model_params=model_params):
return torch.float16
for dt in supported_dtypes:
if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
if torch.float16 in supported_dtypes:
return torch.float16
if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params):
if torch.bfloat16 in supported_dtypes:
return torch.bfloat16
for dt in supported_dtypes:
if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params, manual_cast=True):
if torch.float16 in supported_dtypes:
return torch.float16
if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params, manual_cast=True):
if torch.bfloat16 in supported_dtypes:
return torch.bfloat16
return torch.float32
# None means no manual cast
def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
if weight_dtype == torch.float32 or weight_dtype == torch.float64:
return None
fp16_supported = should_use_fp16(inference_device, prioritize_performance=False)
if fp16_supported and weight_dtype == torch.float16:
return None
bf16_supported = should_use_bf16(inference_device)
if bf16_supported and weight_dtype == torch.bfloat16:
return None
fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
if PRIORITIZE_FP16 and fp16_supported and torch.float16 in supported_dtypes:
return torch.float16
for dt in supported_dtypes:
if dt == torch.float16 and fp16_supported:
return torch.float16
if dt == torch.bfloat16 and bf16_supported:
return torch.bfloat16
return torch.float32
def text_encoder_offload_device():
if args.gpu_only:
return get_torch_device()