forked from comfyanonymous/ComfyUI
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_management.py
957 lines (781 loc) · 29.4 KB
/
model_management.py
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
import psutil
import logging
from enum import Enum
from comfy.cli_args import args
import torch
import sys
import platform
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
lowvram_available = True
xpu_available = False
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:
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
if torch.xpu.is_available():
xpu_available = True
except:
pass
try:
if torch.backends.mps.is_available():
cpu_state = CPUState.MPS
import torch.mps
except:
pass
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 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())
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_torch = mem_reserved
mem_total = torch.xpu.get_device_properties(dev).total_memory
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
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.__version__))
except:
pass
try:
OOM_EXCEPTION = torch.cuda.OutOfMemoryError
except:
OOM_EXCEPTION = Exception
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
ENABLE_PYTORCH_ATTENTION = False
if args.use_pytorch_cross_attention:
ENABLE_PYTORCH_ATTENTION = True
XFORMERS_IS_AVAILABLE = False
VAE_DTYPES = [torch.float32]
try:
if is_nvidia():
torch_version = torch.version.__version__
if int(torch_version[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 torch.cuda.is_bf16_supported() and torch.cuda.get_device_properties(torch.cuda.current_device()).major >= 8:
VAE_DTYPES = [torch.bfloat16] + VAE_DTYPES
if is_intel_xpu():
if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
ENABLE_PYTORCH_ATTENTION = True
except:
pass
if is_intel_xpu():
VAE_DTYPES = [torch.bfloat16] + VAE_DTYPES
if args.cpu_vae:
VAE_DTYPES = [torch.float32]
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)
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
FORCE_FP16 = False
if args.force_fp32:
logging.info("Forcing FP32, if this improves things please report it.")
FORCE_FP32 = True
if args.force_fp16:
logging.info("Forcing FP16.")
FORCE_FP16 = 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)
else:
return "{}".format(device.type)
elif is_intel_xpu():
return "{} {}".format(device, torch.xpu.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.nelement() * t.element_size()
return module_mem
class LoadedModel:
def __init__(self, model):
self.model = model
self.device = model.load_device
self.weights_loaded = False
self.real_model = None
self.currently_used = True
def model_memory(self):
return self.model.model_size()
def model_memory_required(self, device):
if device == self.model.current_device:
return 0
else:
return self.model_memory()
def model_load(self, lowvram_model_memory=0, force_patch_weights=False):
patch_model_to = self.device
self.model.model_patches_to(self.device)
self.model.model_patches_to(self.model.model_dtype())
load_weights = not self.weights_loaded
try:
if lowvram_model_memory > 0 and load_weights:
self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights)
else:
self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights)
except Exception as e:
self.model.unpatch_model(self.model.offload_device)
self.model_unload()
raise e
if is_intel_xpu() and not args.disable_ipex_optimize:
self.real_model = ipex.optimize(self.real_model.eval(), graph_mode=True, concat_linear=True)
self.weights_loaded = True
return self.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, unpatch_weights=True):
self.model.unpatch_model(self.model.offload_device, unpatch_weights=unpatch_weights)
self.model.model_patches_to(self.model.offload_device)
self.weights_loaded = self.weights_loaded and not unpatch_weights
self.real_model = None
def __eq__(self, other):
return self.model is other.model
def minimum_inference_memory():
return (1024 * 1024 * 1024)
def unload_model_clones(model, unload_weights_only=True, force_unload=True):
to_unload = []
for i in range(len(current_loaded_models)):
if model.is_clone(current_loaded_models[i].model):
to_unload = [i] + to_unload
if len(to_unload) == 0:
return True
same_weights = 0
for i in to_unload:
if model.clone_has_same_weights(current_loaded_models[i].model):
same_weights += 1
if same_weights == len(to_unload):
unload_weight = False
else:
unload_weight = True
if not force_unload:
if unload_weights_only and unload_weight == False:
return None
for i in to_unload:
logging.debug("unload clone {} {}".format(i, unload_weight))
current_loaded_models.pop(i).model_unload(unpatch_weights=unload_weight)
return unload_weight
def free_memory(memory_required, device, keep_loaded=[]):
unloaded_model = []
can_unload = []
for i in range(len(current_loaded_models) -1, -1, -1):
shift_model = current_loaded_models[i]
if shift_model.device == device:
if shift_model not in keep_loaded:
can_unload.append((sys.getrefcount(shift_model.model), shift_model.model_memory(), i))
shift_model.currently_used = False
for x in sorted(can_unload):
i = x[-1]
if not DISABLE_SMART_MEMORY:
if get_free_memory(device) > memory_required:
break
current_loaded_models[i].model_unload()
unloaded_model.append(i)
for i in sorted(unloaded_model, reverse=True):
current_loaded_models.pop(i)
if len(unloaded_model) > 0:
soft_empty_cache()
else:
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()
def load_models_gpu(models, memory_required=0, force_patch_weights=False):
global vram_state
inference_memory = minimum_inference_memory()
extra_mem = max(inference_memory, memory_required)
models = set(models)
models_to_load = []
models_already_loaded = []
for x in models:
loaded_model = LoadedModel(x)
loaded = None
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]
if loaded.should_reload_model(force_patch_weights=force_patch_weights): #TODO: cleanup this model reload logic
current_loaded_models.pop(loaded_model_index).model_unload(unpatch_weights=True)
loaded = None
else:
loaded.currently_used = True
models_already_loaded.append(loaded)
if loaded is None:
if hasattr(x, "model"):
logging.info(f"Requested to load {x.model.__class__.__name__}")
models_to_load.append(loaded_model)
if len(models_to_load) == 0:
devs = set(map(lambda a: a.device, models_already_loaded))
for d in devs:
if d != torch.device("cpu"):
free_memory(extra_mem, d, models_already_loaded)
return
logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}")
total_memory_required = {}
for loaded_model in models_to_load:
if unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) == True:#unload clones where the weights are different
total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
for device in total_memory_required:
if device != torch.device("cpu"):
free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded)
for loaded_model in models_to_load:
weights_unloaded = unload_model_clones(loaded_model.model, unload_weights_only=False, force_unload=False) #unload the rest of the clones where the weights can stay loaded
if weights_unloaded is not None:
loaded_model.weights_loaded = not weights_unloaded
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):
model_size = loaded_model.model_memory_required(torch_dev)
current_free_mem = get_free_memory(torch_dev)
lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
if model_size <= (current_free_mem - inference_memory): #only switch to lowvram if really necessary
lowvram_model_memory = 0
if vram_set_state == VRAMState.NO_VRAM:
lowvram_model_memory = 64 * 1024 * 1024
cur_loaded_model = 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(keep_clone_weights_loaded=False):
to_delete = []
for i in range(len(current_loaded_models)):
if sys.getrefcount(current_loaded_models[i].model) <= 2:
if not keep_clone_weights_loaded:
to_delete = [i] + to_delete
#TODO: find a less fragile way to do this.
elif sys.getrefcount(current_loaded_models[i].real_model) <= 3: #references from .real_model + the .model
to_delete = [i] + to_delete
for i in to_delete:
x = current_loaded_models.pop(i)
x.model_unload()
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):
torch_dev = get_torch_device()
if vram_state == VRAMState.HIGH_VRAM:
return torch_dev
cpu_dev = torch.device("cpu")
if DISABLE_SMART_MEMORY:
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 unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
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 should_use_fp16(device=device, model_params=model_params, manual_cast=True):
if torch.float16 in supported_dtypes:
return torch.float16
if 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:
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
if fp16_supported and torch.float16 in supported_dtypes:
return torch.float16
elif bf16_supported and torch.bfloat16 in supported_dtypes:
return torch.bfloat16
else:
return torch.float32
def text_encoder_offload_device():
if args.gpu_only:
return get_torch_device()
else:
return torch.device("cpu")
def text_encoder_device():
if args.gpu_only:
return get_torch_device()
elif vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.NORMAL_VRAM:
if should_use_fp16(prioritize_performance=False):
return get_torch_device()
else:
return torch.device("cpu")
else:
return torch.device("cpu")
def text_encoder_dtype(device=None):
if args.fp8_e4m3fn_text_enc:
return torch.float8_e4m3fn
elif args.fp8_e5m2_text_enc:
return torch.float8_e5m2
elif args.fp16_text_enc:
return torch.float16
elif args.fp32_text_enc:
return torch.float32
if is_device_cpu(device):
return torch.float16
return torch.float16
def intermediate_device():
if args.gpu_only:
return get_torch_device()
else:
return torch.device("cpu")
def vae_device():
if args.cpu_vae:
return torch.device("cpu")
return get_torch_device()
def vae_offload_device():
if args.gpu_only:
return get_torch_device()
else:
return torch.device("cpu")
def vae_dtype(device=None, allowed_dtypes=[]):
global VAE_DTYPES
if args.fp16_vae:
return torch.float16
elif args.bf16_vae:
return torch.bfloat16
elif args.fp32_vae:
return torch.float32
for d in allowed_dtypes:
if d == torch.float16 and should_use_fp16(device, prioritize_performance=False):
return d
if d in VAE_DTYPES:
return d
return VAE_DTYPES[0]
def get_autocast_device(dev):
if hasattr(dev, 'type'):
return dev.type
return "cuda"
def supports_dtype(device, dtype): #TODO
if dtype == torch.float32:
return True
if is_device_cpu(device):
return False
if dtype == torch.float16:
return True
if dtype == torch.bfloat16:
return True
return False
def supports_cast(device, dtype): #TODO
if dtype == torch.float32:
return True
if dtype == torch.float16:
return True
if is_device_mps(device):
return False
if directml_enabled: #TODO: test this
return False
if dtype == torch.bfloat16:
return True
if dtype == torch.float8_e4m3fn:
return True
if dtype == torch.float8_e5m2:
return True
return False
def device_supports_non_blocking(device):
if is_device_mps(device):
return False #pytorch bug? mps doesn't support non blocking
if is_intel_xpu():
return False
if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
return False
if directml_enabled:
return False
return True
def device_should_use_non_blocking(device):
if not device_supports_non_blocking(device):
return False
return False
# return True #TODO: figure out why this causes memory issues on Nvidia and possibly others
def force_channels_last():
if args.force_channels_last:
return True
#TODO
return False
def cast_to_device(tensor, device, dtype, copy=False):
device_supports_cast = False
if tensor.dtype == torch.float32 or tensor.dtype == torch.float16:
device_supports_cast = True
elif tensor.dtype == torch.bfloat16:
if hasattr(device, 'type') and device.type.startswith("cuda"):
device_supports_cast = True
elif is_intel_xpu():
device_supports_cast = True
non_blocking = device_should_use_non_blocking(device)
if device_supports_cast:
if copy:
if tensor.device == device:
return tensor.to(dtype, copy=copy, non_blocking=non_blocking)
return tensor.to(device, copy=copy, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking)
else:
return tensor.to(device, non_blocking=non_blocking).to(dtype, non_blocking=non_blocking)
else:
return tensor.to(device, dtype, copy=copy, non_blocking=non_blocking)
def xformers_enabled():
global directml_enabled
global cpu_state
if cpu_state != CPUState.GPU:
return False
if is_intel_xpu():
return False
if directml_enabled:
return False
return XFORMERS_IS_AVAILABLE
def xformers_enabled_vae():
enabled = xformers_enabled()
if not enabled:
return False
return XFORMERS_ENABLED_VAE
def pytorch_attention_enabled():
global ENABLE_PYTORCH_ATTENTION
return ENABLE_PYTORCH_ATTENTION
def pytorch_attention_flash_attention():
global ENABLE_PYTORCH_ATTENTION
if ENABLE_PYTORCH_ATTENTION:
#TODO: more reliable way of checking for flash attention?
if is_nvidia(): #pytorch flash attention only works on Nvidia
return True
if is_intel_xpu():
return True
return False
def force_upcast_attention_dtype():
upcast = args.force_upcast_attention
try:
if platform.mac_ver()[0] in ['14.5']: #black image bug on OSX Sonoma 14.5
upcast = True
except:
pass
if upcast:
return torch.float32
else:
return None
def get_free_memory(dev=None, torch_free_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_free_total = psutil.virtual_memory().available
mem_free_torch = mem_free_total
else:
if directml_enabled:
mem_free_total = 1024 * 1024 * 1024 #TODO
mem_free_torch = mem_free_total
elif is_intel_xpu():
stats = torch.xpu.memory_stats(dev)
mem_active = stats['active_bytes.all.current']
mem_reserved = stats['reserved_bytes.all.current']
mem_free_torch = mem_reserved - mem_active
mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved
mem_free_total = mem_free_xpu + mem_free_torch
else:
stats = torch.cuda.memory_stats(dev)
mem_active = stats['active_bytes.all.current']
mem_reserved = stats['reserved_bytes.all.current']
mem_free_cuda, _ = torch.cuda.mem_get_info(dev)
mem_free_torch = mem_reserved - mem_active
mem_free_total = mem_free_cuda + mem_free_torch
if torch_free_too:
return (mem_free_total, mem_free_torch)
else:
return mem_free_total
def cpu_mode():
global cpu_state
return cpu_state == CPUState.CPU
def mps_mode():
global cpu_state
return cpu_state == CPUState.MPS
def is_device_type(device, type):
if hasattr(device, 'type'):
if (device.type == type):
return True
return False
def is_device_cpu(device):
return is_device_type(device, 'cpu')
def is_device_mps(device):
return is_device_type(device, 'mps')
def is_device_cuda(device):
return is_device_type(device, 'cuda')
def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
global directml_enabled
if device is not None:
if is_device_cpu(device):
return False
if FORCE_FP16:
return True
if device is not None:
if is_device_mps(device):
return True
if FORCE_FP32:
return False
if directml_enabled:
return False
if mps_mode():
return True
if cpu_mode():
return False
if is_intel_xpu():
return True
if torch.version.hip:
return True
props = torch.cuda.get_device_properties("cuda")
if props.major >= 8:
return True
if props.major < 6:
return False
fp16_works = False
#FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
#when the model doesn't actually fit on the card
#TODO: actually test if GP106 and others have the same type of behavior
nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
for x in nvidia_10_series:
if x in props.name.lower():
fp16_works = True
if fp16_works or manual_cast:
free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
if (not prioritize_performance) or model_params * 4 > free_model_memory:
return True
if props.major < 7:
return False
#FP16 is just broken on these cards
nvidia_16_series = ["1660", "1650", "1630", "T500", "T550", "T600", "MX550", "MX450", "CMP 30HX", "T2000", "T1000", "T1200"]
for x in nvidia_16_series:
if x in props.name:
return False
return True
def should_use_bf16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
if device is not None:
if is_device_cpu(device): #TODO ? bf16 works on CPU but is extremely slow
return False
if device is not None: #TODO not sure about mps bf16 support
if is_device_mps(device):
return False
if FORCE_FP32:
return False
if directml_enabled:
return False
if cpu_mode() or mps_mode():
return False
if is_intel_xpu():
return True
if device is None:
device = torch.device("cuda")
props = torch.cuda.get_device_properties(device)
if props.major >= 8:
return True
bf16_works = torch.cuda.is_bf16_supported()
if bf16_works or manual_cast:
free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
if (not prioritize_performance) or model_params * 4 > free_model_memory:
return True
return False
def soft_empty_cache(force=False):
global cpu_state
if cpu_state == CPUState.MPS:
torch.mps.empty_cache()
elif is_intel_xpu():
torch.xpu.empty_cache()
elif torch.cuda.is_available():
if force or is_nvidia(): #This seems to make things worse on ROCm so I only do it for cuda
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
def unload_all_models():
free_memory(1e30, get_torch_device())
def resolve_lowvram_weight(weight, model, key): #TODO: remove
print("WARNING: The comfy.model_management.resolve_lowvram_weight function will be removed soon, please stop using it.")
return weight
#TODO: might be cleaner to put this somewhere else
import threading
class InterruptProcessingException(Exception):
pass
interrupt_processing_mutex = threading.RLock()
interrupt_processing = False
def interrupt_current_processing(value=True):
global interrupt_processing
global interrupt_processing_mutex
with interrupt_processing_mutex:
interrupt_processing = value
def processing_interrupted():
global interrupt_processing
global interrupt_processing_mutex
with interrupt_processing_mutex:
return interrupt_processing
def throw_exception_if_processing_interrupted():
global interrupt_processing
global interrupt_processing_mutex
with interrupt_processing_mutex:
if interrupt_processing:
interrupt_processing = False
raise InterruptProcessingException()