-
Notifications
You must be signed in to change notification settings - Fork 114
/
Copy pathnodes.py
1118 lines (967 loc) · 50.8 KB
/
nodes.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
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
import os
import torch
import json
import gc
from .utils import log, print_memory
from diffusers.video_processor import VideoProcessor
from .hyvideo.constants import PROMPT_TEMPLATE
from .hyvideo.text_encoder import TextEncoder
from .hyvideo.utils.data_utils import align_to
from .hyvideo.modules.posemb_layers import get_nd_rotary_pos_embed
from .hyvideo.diffusion.schedulers import FlowMatchDiscreteScheduler
from .hyvideo.diffusion.pipelines import HunyuanVideoPipeline
from .hyvideo.vae.autoencoder_kl_causal_3d import AutoencoderKLCausal3D
from .hyvideo.modules.models import HYVideoDiffusionTransformer
from accelerate import init_empty_weights
from accelerate.utils import set_module_tensor_to_device
import folder_paths
import comfy.model_management as mm
from comfy.utils import load_torch_file
import comfy.model_base
import comfy.latent_formats
script_directory = os.path.dirname(os.path.abspath(__file__))
def get_rotary_pos_embed(transformer, video_length, height, width):
target_ndim = 3
ndim = 5 - 2
rope_theta = 225
patch_size = transformer.patch_size
rope_dim_list = transformer.rope_dim_list
hidden_size = transformer.hidden_size
heads_num = transformer.heads_num
head_dim = hidden_size // heads_num
# 884
latents_size = [(video_length - 1) // 4 + 1, height // 8, width // 8]
if isinstance(patch_size, int):
assert all(s % patch_size == 0 for s in latents_size), (
f"Latent size(last {ndim} dimensions) should be divisible by patch size({patch_size}), "
f"but got {latents_size}."
)
rope_sizes = [s // patch_size for s in latents_size]
elif isinstance(patch_size, list):
assert all(
s % patch_size[idx] == 0
for idx, s in enumerate(latents_size)
), (
f"Latent size(last {ndim} dimensions) should be divisible by patch size({patch_size}), "
f"but got {latents_size}."
)
rope_sizes = [
s // patch_size[idx] for idx, s in enumerate(latents_size)
]
if len(rope_sizes) != target_ndim:
rope_sizes = [1] * (target_ndim - len(rope_sizes)) + rope_sizes # time axis
if rope_dim_list is None:
rope_dim_list = [head_dim // target_ndim for _ in range(target_ndim)]
assert (
sum(rope_dim_list) == head_dim
), "sum(rope_dim_list) should equal to head_dim of attention layer"
freqs_cos, freqs_sin = get_nd_rotary_pos_embed(
rope_dim_list,
rope_sizes,
theta=rope_theta,
use_real=True,
theta_rescale_factor=1,
)
return freqs_cos, freqs_sin
class HyVideoBlockSwap:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"double_blocks_to_swap": ("INT", {"default": 20, "min": 0, "max": 20, "step": 1, "tooltip": "Number of double blocks to swap"}),
"single_blocks_to_swap": ("INT", {"default": 0, "min": 0, "max": 40, "step": 1, "tooltip": "Number of single blocks to swap"}),
"offload_txt_in": ("BOOLEAN", {"default": False, "tooltip": "Offload txt_in layer"}),
"offload_img_in": ("BOOLEAN", {"default": False, "tooltip": "Offload img_in layer"}),
},
}
RETURN_TYPES = ("BLOCKSWAPARGS",)
RETURN_NAMES = ("block_swap_args",)
FUNCTION = "setargs"
CATEGORY = "HunyuanVideoWrapper"
DESCRIPTION = "Settings for block swapping, reduces VRAM use by swapping blocks to CPU memory"
def setargs(self, **kwargs):
return (kwargs, )
class HyVideoSTG:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"stg_mode": (["STG-A", "STG-R"],),
"stg_block_idx": ("INT", {"default": 0, "min": -1, "max": 39, "step": 1, "tooltip": "Block index to apply STG"}),
"stg_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01, "tooltip": "Recommended values are ≤2.0"}),
"stg_start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01, "tooltip": "Start percentage of the steps to apply STG"}),
"stg_end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01, "tooltip": "End percentage of the steps to apply STG"}),
},
}
RETURN_TYPES = ("STGARGS",)
RETURN_NAMES = ("stg_args",)
FUNCTION = "setargs"
CATEGORY = "HunyuanVideoWrapper"
DESCRIPTION = "Spatio Temporal Guidance, https://github.com/junhahyung/STGuidance"
def setargs(self, **kwargs):
return (kwargs, )
class HyVideoModel(comfy.model_base.BaseModel):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.pipeline = {}
def __getitem__(self, k):
return self.pipeline[k]
def __setitem__(self, k, v):
self.pipeline[k] = v
class HyVideoModelConfig:
def __init__(self, dtype):
self.unet_config = {}
self.unet_extra_config = {}
self.latent_format = comfy.latent_formats.LatentFormat()
self.latent_format.latent_channels = 16
self.manual_cast_dtype = dtype
self.sampling_settings = {"multiplier": 1.0}
# Don't know what this is. Value taken from ComfyUI Mochi model.
self.memory_usage_factor = 2.0
# denoiser is handled by extension
self.unet_config["disable_unet_model_creation"] = True
#region Model loading
class HyVideoModelLoader:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": (folder_paths.get_filename_list("diffusion_models"), {"tooltip": "These models are loaded from the 'ComfyUI/models/diffusion_models' -folder",}),
"base_precision": (["fp16", "fp32", "bf16"], {"default": "bf16"}),
"quantization": (['disabled', 'fp8_e4m3fn', 'fp8_e4m3fn_fast', 'torchao_fp8dq', "torchao_fp8dqrow", "torchao_int8dq", "torchao_fp6", "torchao_int4", "torchao_int8"], {"default": 'disabled', "tooltip": "optional quantization method"}),
"load_device": (["main_device", "offload_device"], {"default": "main_device"}),
},
"optional": {
"attention_mode": ([
"sdpa",
"flash_attn_varlen",
"sageattn_varlen",
"comfy",
], {"default": "flash_attn"}),
"compile_args": ("COMPILEARGS", ),
"block_swap_args": ("BLOCKSWAPARGS", ),
}
}
RETURN_TYPES = ("MODEL",)
RETURN_NAMES = ("model", )
FUNCTION = "loadmodel"
CATEGORY = "HunyuanVideoWrapper"
def loadmodel(self, model, base_precision, load_device, quantization,
compile_args=None, attention_mode="sdpa", block_swap_args=None):
transformer = None
manual_offloading = True
if "sage" in attention_mode:
try:
from sageattention import sageattn_varlen
except Exception as e:
raise ValueError(f"Can't import SageAttention: {str(e)}")
device = mm.get_torch_device()
offload_device = mm.unet_offload_device()
manual_offloading = True
transformer_load_device = device if load_device == "main_device" else offload_device
mm.soft_empty_cache()
base_dtype = {"fp8_e4m3fn": torch.float8_e4m3fn, "fp8_e4m3fn_fast": torch.float8_e4m3fn, "bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[base_precision]
model_path = folder_paths.get_full_path_or_raise("diffusion_models", model)
sd = load_torch_file(model_path, device=offload_device)
in_channels = out_channels = 16
factor_kwargs = {"device": transformer_load_device, "dtype": base_dtype}
HUNYUAN_VIDEO_CONFIG = {
"mm_double_blocks_depth": 20,
"mm_single_blocks_depth": 40,
"rope_dim_list": [16, 56, 56],
"hidden_size": 3072,
"heads_num": 24,
"mlp_width_ratio": 4,
"guidance_embed": True,
}
with init_empty_weights():
transformer = HYVideoDiffusionTransformer(
in_channels=in_channels,
out_channels=out_channels,
attention_mode=attention_mode,
main_device=device,
offload_device=offload_device,
**HUNYUAN_VIDEO_CONFIG,
**factor_kwargs
)
transformer.eval()
if "torchao" in quantization:
try:
from torchao.quantization import (
quantize_,
fpx_weight_only,
float8_dynamic_activation_float8_weight,
int8_dynamic_activation_int8_weight,
int8_weight_only,
int4_weight_only
)
except:
raise ImportError("torchao is not installed, please install torchao to use fp8dq")
# def filter_fn(module: nn.Module, fqn: str) -> bool:
# target_submodules = {'attn1', 'ff'} # avoid norm layers, 1.5 at least won't work with quantized norm1 #todo: test other models
# if any(sub in fqn for sub in target_submodules):
# return isinstance(module, nn.Linear)
# return False
if "fp6" in quantization:
quant_func = fpx_weight_only(3, 2)
elif "int4" in quantization:
quant_func = int4_weight_only()
elif "int8" in quantization:
quant_func = int8_weight_only()
elif "fp8dq" in quantization:
quant_func = float8_dynamic_activation_float8_weight()
elif 'fp8dqrow' in quantization:
from torchao.quantization.quant_api import PerRow
quant_func = float8_dynamic_activation_float8_weight(granularity=PerRow())
elif 'int8dq' in quantization:
quant_func = int8_dynamic_activation_int8_weight()
log.info(f"Quantizing model with {quant_func}")
for i, block in enumerate(transformer.single_blocks):
log.info(f"Quantizing single_block {i}")
for name, _ in block.named_parameters(prefix=f"single_blocks.{i}"):
#print(f"Parameter name: {name}")
set_module_tensor_to_device(transformer, name, device=transformer_load_device, dtype=base_dtype, value=sd[name])
if compile_args is not None:
transformer.single_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
quantize_(block, quant_func)
print(block)
block.to(offload_device)
for i, block in enumerate(transformer.double_blocks):
log.info(f"Quantizing double_block {i}")
for name, _ in block.named_parameters(prefix=f"double_blocks.{i}"):
#print(f"Parameter name: {name}")
set_module_tensor_to_device(transformer, name, device=transformer_load_device, dtype=base_dtype, value=sd[name])
if compile_args is not None:
transformer.double_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
quantize_(block, quant_func)
for name, param in transformer.named_parameters():
if "single_blocks" not in name and "double_blocks" not in name:
set_module_tensor_to_device(transformer, name, device=transformer_load_device, dtype=base_dtype, value=sd[name])
manual_offloading = False # to disable manual .to(device) calls
log.info(f"Quantized transformer blocks to {quantization}")
for name, param in transformer.named_parameters():
print(name, param.dtype)
#param.data = param.data.to(self.vae_dtype).to(device)
else:
log.info("Using accelerate to load and assign model weights to device...")
if quantization == "fp8_e4m3fn" or quantization == "fp8_e4m3fn_fast":
dtype = torch.float8_e4m3fn
else:
dtype = base_dtype
params_to_keep = {"norm", "bias", "time_in", "vector_in", "guidance_in", "txt_in", "img_in"}
for name, param in transformer.named_parameters():
dtype_to_use = base_dtype if any(keyword in name for keyword in params_to_keep) else dtype
set_module_tensor_to_device(transformer, name, device=transformer_load_device, dtype=dtype_to_use, value=sd[name])
if quantization == "fp8_e4m3fn_fast":
from .fp8_optimization import convert_fp8_linear
convert_fp8_linear(transformer, base_dtype, params_to_keep=params_to_keep)
#compile
if compile_args is not None:
torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
if compile_args["compile_single_blocks"]:
for i, block in enumerate(transformer.single_blocks):
transformer.single_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
if compile_args["compile_double_blocks"]:
for i, block in enumerate(transformer.double_blocks):
transformer.double_blocks[i] = torch.compile(block, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
if compile_args["compile_txt_in"]:
transformer.txt_in = torch.compile(transformer.txt_in, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
if compile_args["compile_vector_in"]:
transformer.vector_in = torch.compile(transformer.vector_in, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
if compile_args["compile_final_layer"]:
transformer.final_layer = torch.compile(transformer.final_layer, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
del sd
mm.soft_empty_cache()
scheduler = FlowMatchDiscreteScheduler(
shift=9.0,
reverse=True,
solver="euler",
)
pipe = HunyuanVideoPipeline(
transformer=transformer,
scheduler=scheduler,
progress_bar_config=None,
base_dtype=base_dtype
)
comfy_model = HyVideoModel(
HyVideoModelConfig(base_dtype),
model_type=comfy.model_base.ModelType.FLOW,
device=mm.get_torch_device(),
)
comfy_model["pipe"] = pipe
comfy_model["dtype"] = base_dtype
comfy_model["base_path"] = model_path
comfy_model["model_name"] = model
comfy_model["manual_offloading"] = manual_offloading
comfy_model["quantization"] = "disabled"
comfy_model["block_swap_args"] = block_swap_args
comfy_model.diffusion_model = transformer
patcher = comfy.model_patcher.ModelPatcher(comfy_model, device, offload_device)
return (patcher,)
#region load VAE
class HyVideoVAELoader:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model_name": (folder_paths.get_filename_list("vae"), {"tooltip": "These models are loaded from 'ComfyUI/models/vae'"}),
},
"optional": {
"precision": (["fp16", "fp32", "bf16"],
{"default": "bf16"}
),
"compile_args":("COMPILEARGS", ),
}
}
RETURN_TYPES = ("VAE",)
RETURN_NAMES = ("vae", )
FUNCTION = "loadmodel"
CATEGORY = "HunyuanVideoWrapper"
DESCRIPTION = "Loads Hunyuan VAE model from 'ComfyUI/models/vae'"
def loadmodel(self, model_name, precision, compile_args=None):
device = mm.get_torch_device()
offload_device = mm.unet_offload_device()
dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[precision]
with open(os.path.join(script_directory, 'configs', 'hy_vae_config.json')) as f:
vae_config = json.load(f)
model_path = folder_paths.get_full_path("vae", model_name)
vae_sd = load_torch_file(model_path)
vae = AutoencoderKLCausal3D.from_config(vae_config)
vae.load_state_dict(vae_sd)
del vae_sd
vae.requires_grad_(False)
vae.eval()
vae.to(device = device, dtype = dtype)
#compile
if compile_args is not None:
torch._dynamo.config.cache_size_limit = compile_args["dynamo_cache_size_limit"]
vae = torch.compile(vae, fullgraph=compile_args["fullgraph"], dynamic=compile_args["dynamic"], backend=compile_args["backend"], mode=compile_args["mode"])
return (vae,)
class HyVideoTorchCompileSettings:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"backend": (["inductor","cudagraphs"], {"default": "inductor"}),
"fullgraph": ("BOOLEAN", {"default": False, "tooltip": "Enable full graph mode"}),
"mode": (["default", "max-autotune", "max-autotune-no-cudagraphs", "reduce-overhead"], {"default": "default"}),
"dynamic": ("BOOLEAN", {"default": False, "tooltip": "Enable dynamic mode"}),
"dynamo_cache_size_limit": ("INT", {"default": 64, "min": 0, "max": 1024, "step": 1, "tooltip": "torch._dynamo.config.cache_size_limit"}),
"compile_single_blocks": ("BOOLEAN", {"default": True, "tooltip": "Compile single blocks"}),
"compile_double_blocks": ("BOOLEAN", {"default": True, "tooltip": "Compile double blocks"}),
"compile_txt_in": ("BOOLEAN", {"default": False, "tooltip": "Compile txt_in layers"}),
"compile_vector_in": ("BOOLEAN", {"default": False, "tooltip": "Compile vector_in layers"}),
"compile_final_layer": ("BOOLEAN", {"default": False, "tooltip": "Compile final layer"}),
},
}
RETURN_TYPES = ("COMPILEARGS",)
RETURN_NAMES = ("torch_compile_args",)
FUNCTION = "loadmodel"
CATEGORY = "HunyuanVideoWrapper"
DESCRIPTION = "torch.compile settings, when connected to the model loader, torch.compile of the selected layers is attempted. Requires Triton and torch 2.5.0 is recommended"
def loadmodel(self, backend, fullgraph, mode, dynamic, dynamo_cache_size_limit, compile_single_blocks, compile_double_blocks, compile_txt_in, compile_vector_in, compile_final_layer):
compile_args = {
"backend": backend,
"fullgraph": fullgraph,
"mode": mode,
"dynamic": dynamic,
"dynamo_cache_size_limit": dynamo_cache_size_limit,
"compile_single_blocks": compile_single_blocks,
"compile_double_blocks": compile_double_blocks,
"compile_txt_in": compile_txt_in,
"compile_vector_in": compile_vector_in,
"compile_final_layer": compile_final_layer
}
return (compile_args, )
#region TextEncode
class DownloadAndLoadHyVideoTextEncoder:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"llm_model": (["Kijai/llava-llama-3-8b-text-encoder-tokenizer",],),
"clip_model": (["disabled","openai/clip-vit-large-patch14",],),
"precision": (["fp16", "fp32", "bf16"],
{"default": "bf16"}
),
},
"optional": {
"apply_final_norm": ("BOOLEAN", {"default": False}),
"hidden_state_skip_layer": ("INT", {"default": 2}),
"quantization": (['disabled', 'bnb_nf4'], {"default": 'disabled'}),
}
}
RETURN_TYPES = ("HYVIDTEXTENCODER",)
RETURN_NAMES = ("hyvid_text_encoder", )
FUNCTION = "loadmodel"
CATEGORY = "HunyuanVideoWrapper"
DESCRIPTION = "Loads Hunyuan text_encoder model from 'ComfyUI/models/LLM'"
def loadmodel(self, llm_model, clip_model, precision, apply_final_norm=False, hidden_state_skip_layer=2, quantization="disabled"):
device = mm.get_torch_device()
offload_device = mm.unet_offload_device()
dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[precision]
if quantization == "bnb_nf4":
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16
)
else:
quantization_config = None
if clip_model != "disabled":
clip_model_path = os.path.join(folder_paths.models_dir, "clip", "clip-vit-large-patch14")
if not os.path.exists(clip_model_path):
log.info(f"Downloading clip model to: {clip_model_path}")
from huggingface_hub import snapshot_download
snapshot_download(
repo_id=clip_model,
ignore_patterns=["*.msgpack", "*.bin", "*.h5"],
local_dir=clip_model_path,
local_dir_use_symlinks=False,
)
text_encoder_2 = TextEncoder(
text_encoder_path=clip_model_path,
text_encoder_type="clipL",
max_length=77,
text_encoder_precision=precision,
tokenizer_type="clipL",
logger=log,
device=device,
)
else:
text_encoder_2 = None
download_path = os.path.join(folder_paths.models_dir,"LLM")
base_path = os.path.join(download_path, (llm_model.split("/")[-1]))
if not os.path.exists(base_path):
log.info(f"Downloading model to: {base_path}")
from huggingface_hub import snapshot_download
snapshot_download(
repo_id=llm_model,
local_dir=base_path,
local_dir_use_symlinks=False,
)
text_encoder = TextEncoder(
text_encoder_path=base_path,
text_encoder_type="llm",
max_length=256,
text_encoder_precision=precision,
tokenizer_type="llm",
hidden_state_skip_layer=hidden_state_skip_layer,
apply_final_norm=apply_final_norm,
logger=log,
device=device,
dtype=dtype,
quantization_config=quantization_config
)
hyvid_text_encoders = {
"text_encoder": text_encoder,
"text_encoder_2": text_encoder_2,
}
return (hyvid_text_encoders,)
class HyVideoCustomPromptTemplate:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"custom_prompt_template": ("STRING", {"default": f"{PROMPT_TEMPLATE['dit-llm-encode-video']['template']}", "multiline": True}),
"crop_start": ("INT", {"default": PROMPT_TEMPLATE['dit-llm-encode-video']["crop_start"], "tooltip": "To cropt the system prompt"}),
},
}
RETURN_TYPES = ("PROMPT_TEMPLATE", )
RETURN_NAMES = ("hyvid_prompt_template",)
FUNCTION = "process"
CATEGORY = "HunyuanVideoWrapper"
def process(self, custom_prompt_template, crop_start):
prompt_template_dict = {
"template": custom_prompt_template,
"crop_start": crop_start,
}
return (prompt_template_dict,)
class HyVideoTextEncode:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"text_encoders": ("HYVIDTEXTENCODER",),
"prompt": ("STRING", {"default": "", "multiline": True} ),
#"negative_prompt": ("STRING", {"default": "", "multiline": True}),
},
"optional": {
"force_offload": ("BOOLEAN", {"default": True}),
"prompt_template": (["video", "image", "custom", "disabled"], {"default": "video", "tooltip": "Use the default prompt templates for the llm text encoder"}),
"custom_prompt_template": ("PROMPT_TEMPLATE", {"default": PROMPT_TEMPLATE["dit-llm-encode-video"], "multiline": True}),
"clip_l": ("CLIP", {"tooltip": "Use comfy clip model instead, in this case the text encoder loader's clip_l should be disabled"}),
}
}
RETURN_TYPES = ("HYVIDEMBEDS", )
RETURN_NAMES = ("hyvid_embeds",)
FUNCTION = "process"
CATEGORY = "HunyuanVideoWrapper"
def process(self, text_encoders, prompt, force_offload=True, prompt_template="video", custom_prompt_template=None, clip_l=None):
device = mm.text_encoder_device()
offload_device = mm.text_encoder_offload_device()
text_encoder_1 = text_encoders["text_encoder"]
if clip_l is None:
text_encoder_2 = text_encoders["text_encoder_2"]
else:
text_encoder_2 = None
negative_prompt = None
if prompt_template != "disabled":
if prompt_template == "custom":
prompt_template_dict = custom_prompt_template
elif prompt_template == "video":
prompt_template_dict = PROMPT_TEMPLATE["dit-llm-encode-video"]
elif prompt_template == "image":
prompt_template_dict = PROMPT_TEMPLATE["dit-llm-encode"]
else:
raise ValueError(f"Invalid prompt_template: {prompt_template_dict}")
assert (
isinstance(prompt_template_dict, dict)
and "template" in prompt_template_dict
), f"`prompt_template` must be a dictionary with a key 'template', got {prompt_template_dict}"
assert "{}" in str(prompt_template_dict["template"]), (
"`prompt_template['template']` must contain a placeholder `{}` for the input text, "
f"got {prompt_template_dict['template']}"
)
else:
prompt_template_dict = None
def encode_prompt(self, prompt, negative_prompt, text_encoder):
batch_size = 1
num_videos_per_prompt = 1
do_classifier_free_guidance = False # not implemented, for now we only have cfg distilled model
text_inputs = text_encoder.text2tokens(prompt, prompt_template=prompt_template_dict)
prompt_outputs = text_encoder.encode(text_inputs, prompt_template=prompt_template_dict, device=device)
prompt_embeds = prompt_outputs.hidden_state
attention_mask = prompt_outputs.attention_mask
log.info(f"{text_encoder.text_encoder_type} prompt attention_mask shape: {attention_mask.shape}, masked tokens: {attention_mask[0].sum().item()}")
if attention_mask is not None:
attention_mask = attention_mask.to(device)
bs_embed, seq_len = attention_mask.shape
attention_mask = attention_mask.repeat(1, num_videos_per_prompt)
attention_mask = attention_mask.view(
bs_embed * num_videos_per_prompt, seq_len
)
prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)
# if prompt_embeds.ndim == 2:
# bs_embed, _ = prompt_embeds.shape
# # duplicate text embeddings for each generation per prompt, using mps friendly method
# prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
# prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, -1)
# else:
# bs_embed, seq_len, _ = prompt_embeds.shape
# # duplicate text embeddings for each generation per prompt, using mps friendly method
# prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
# prompt_embeds = prompt_embeds.view(
# bs_embed * num_videos_per_prompt, seq_len, -1
# )
# get unconditional embeddings for classifier free guidance
# if do_classifier_free_guidance:
# uncond_tokens: List[str]
# if negative_prompt is None:
# uncond_tokens = [""] * batch_size
# elif prompt is not None and type(prompt) is not type(negative_prompt):
# raise TypeError(
# f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
# f" {type(prompt)}."
# )
# elif isinstance(negative_prompt, str):
# uncond_tokens = [negative_prompt]
# elif batch_size != len(negative_prompt):
# raise ValueError(
# f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
# f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
# " the batch size of `prompt`."
# )
# else:
# uncond_tokens = negative_prompt
# # max_length = prompt_embeds.shape[1]
# uncond_input = text_encoder.text2tokens(uncond_tokens, data_type=data_type)
# negative_prompt_outputs = text_encoder.encode(
# uncond_input, data_type=data_type, device=device
# )
# negative_prompt_embeds = negative_prompt_outputs.hidden_state
# negative_attention_mask = negative_prompt_outputs.attention_mask
# if negative_attention_mask is not None:
# negative_attention_mask = negative_attention_mask.to(device)
# _, seq_len = negative_attention_mask.shape
# negative_attention_mask = negative_attention_mask.repeat(
# 1, num_videos_per_prompt
# )
# negative_attention_mask = negative_attention_mask.view(
# batch_size * num_videos_per_prompt, seq_len
# )
# else:
negative_prompt_embeds = None
negative_attention_mask = None
# if do_classifier_free_guidance:
# # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
# seq_len = negative_prompt_embeds.shape[1]
# negative_prompt_embeds = negative_prompt_embeds.to(
# dtype=prompt_embeds_dtype, device=device
# )
# if negative_prompt_embeds.ndim == 2:
# negative_prompt_embeds = negative_prompt_embeds.repeat(
# 1, num_videos_per_prompt
# )
# negative_prompt_embeds = negative_prompt_embeds.view(
# batch_size * num_videos_per_prompt, -1
# )
# else:
# negative_prompt_embeds = negative_prompt_embeds.repeat(
# 1, num_videos_per_prompt, 1
# )
# negative_prompt_embeds = negative_prompt_embeds.view(
# batch_size * num_videos_per_prompt, seq_len, -1
# )
return (
prompt_embeds,
negative_prompt_embeds,
attention_mask,
negative_attention_mask,
)
text_encoder_1.to(device)
prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = encode_prompt(self, prompt, negative_prompt, text_encoder_1)
if force_offload:
text_encoder_1.to(offload_device)
mm.soft_empty_cache()
if text_encoder_2 is not None:
text_encoder_2.to(device)
prompt_embeds_2, negative_prompt_embeds_2, attention_mask_2, negative_attention_mask_2 = encode_prompt(self, prompt, negative_prompt, text_encoder_2)
if force_offload:
text_encoder_2.to(offload_device)
mm.soft_empty_cache()
elif clip_l is not None:
clip_l.cond_stage_model.to(device)
tokens = clip_l.tokenize(prompt, return_word_ids=True)
prompt_embeds_2 = clip_l.encode_from_tokens(tokens, return_pooled=True, return_dict=False)[1]
prompt_embeds_2 = prompt_embeds_2.to(device=device)
negative_prompt_embeds_2, attention_mask_2, negative_attention_mask_2 = None, None, None
if force_offload:
clip_l.cond_stage_model.to(offload_device)
mm.soft_empty_cache()
else:
prompt_embeds_2 = None
negative_prompt_embeds_2 = None
attention_mask_2 = None
negative_attention_mask_2 = None
prompt_embeds_dict = {
"prompt_embeds": prompt_embeds,
"negative_prompt_embeds": negative_prompt_embeds,
"attention_mask": attention_mask,
"negative_attention_mask": negative_attention_mask,
"prompt_embeds_2": prompt_embeds_2,
"negative_prompt_embeds_2": negative_prompt_embeds_2,
"attention_mask_2": attention_mask_2,
"negative_attention_mask_2": negative_attention_mask_2,
}
return (prompt_embeds_dict,)
#region Sampler
class HyVideoSampler:
@classmethod
def INPUT_TYPES(s):
return {
"required": {
"model": ("MODEL",),
"hyvid_embeds": ("HYVIDEMBEDS", ),
"width": ("INT", {"default": 512, "min": 64, "max": 4096, "step": 16}),
"height": ("INT", {"default": 512, "min": 64, "max": 4096, "step": 16}),
"num_frames": ("INT", {"default": 49, "min": 1, "max": 1024, "step": 4}),
"steps": ("INT", {"default": 30, "min": 1}),
"embedded_guidance_scale": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 30.0, "step": 0.01}),
"flow_shift": ("FLOAT", {"default": 9.0, "min": 0.0, "max": 30.0, "step": 0.01}),
"seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
"force_offload": ("BOOLEAN", {"default": True}),
},
"optional": {
"samples": ("LATENT", {"tooltip": "init Latents to use for video2video process"} ),
"denoise_strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
"stg_args": ("STGARGS", ),
}
}
RETURN_TYPES = ("LATENT",)
RETURN_NAMES = ("samples",)
FUNCTION = "process"
CATEGORY = "HunyuanVideoWrapper"
def process(self, model, hyvid_embeds, flow_shift, steps, embedded_guidance_scale, seed, width, height, num_frames, samples=None, denoise_strength=1.0, force_offload=True, stg_args=None):
# We only need this so that LoRA weights get patched.
comfy.model_management.load_models_gpu([model])
model = model.model
device = mm.get_torch_device()
offload_device = mm.unet_offload_device()
dtype = model["dtype"]
transformer = model["pipe"].transformer
if stg_args is not None:
if stg_args["stg_mode"] == "STG-A" and transformer.attention_mode != "sdpa":
raise ValueError(
f"STG-A requires attention_mode to be 'sdpa', but got {transformer.attention_mode}."
)
generator = torch.Generator(device=torch.device("cpu")).manual_seed(seed)
if width <= 0 or height <= 0 or num_frames <= 0:
raise ValueError(
f"`height` and `width` and `video_length` must be positive integers, got height={height}, width={width}, video_length={num_frames}"
)
if (num_frames - 1) % 4 != 0:
raise ValueError(
f"`video_length-1` must be a multiple of 4, got {num_frames}"
)
log.info(
f"Input (height, width, video_length) = ({height}, {width}, {num_frames})"
)
target_height = align_to(height, 16)
target_width = align_to(width, 16)
freqs_cos, freqs_sin = get_rotary_pos_embed(
model["pipe"].transformer, num_frames, target_height, target_width
)
n_tokens = freqs_cos.shape[0]
model["pipe"].scheduler.shift = flow_shift
# autocast_context = torch.autocast(
# mm.get_autocast_device(device), dtype=dtype
# ) if any(q in model["quantization"] for q in ("e4m3fn", "GGUF")) else nullcontext()
#with autocast_context:
if model["block_swap_args"] is not None:
for name, param in transformer.named_parameters():
#print(name, param.data.device)
if "single" not in name and "double" not in name:
param.data = param.data.to(device)
transformer.block_swap(
model["block_swap_args"]["double_blocks_to_swap"] - 1 ,
model["block_swap_args"]["single_blocks_to_swap"] - 1,
offload_txt_in = model["block_swap_args"]["offload_txt_in"],
offload_img_in = model["block_swap_args"]["offload_img_in"],
)
mm.soft_empty_cache()
gc.collect()
elif model["manual_offloading"]:
transformer.to(device)
mm.soft_empty_cache()
gc.collect()
try:
torch.cuda.reset_peak_memory_stats(device)
except:
pass
#for name, param in transformer.named_parameters():
# print(name, param.data.device)
out_latents = model["pipe"](
num_inference_steps=steps,
height = target_height,
width = target_width,
video_length = num_frames,
guidance_scale=1.0,
embedded_guidance_scale=embedded_guidance_scale,
latents=samples["samples"] if samples is not None else None,
denoise_strength=denoise_strength,
prompt_embed_dict=hyvid_embeds,
generator=generator,
freqs_cis=(freqs_cos, freqs_sin),
n_tokens=n_tokens,
stg_mode=stg_args["stg_mode"] if stg_args is not None else None,
stg_block_idx=stg_args["stg_block_idx"] if stg_args is not None else -1,
stg_scale=stg_args["stg_scale"] if stg_args is not None else 0.0,
stg_start_percent=stg_args["stg_start_percent"] if stg_args is not None else 0.0,
stg_end_percent=stg_args["stg_end_percent"] if stg_args is not None else 1.0,
)
print_memory(device)
try:
torch.cuda.reset_peak_memory_stats(device)
except:
pass
if force_offload:
if model["manual_offloading"]:
transformer.to(offload_device)
mm.soft_empty_cache()
gc.collect()
return ({
"samples": out_latents
},)
#region VideoDecode
class HyVideoDecode:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"vae": ("VAE",),
"samples": ("LATENT",),
"enable_vae_tiling": ("BOOLEAN", {"default": True, "tooltip": "Drastically reduces memory use but may introduce seams"}),
"temporal_tiling_sample_size": ("INT", {"default": 16, "min": 4, "max": 256, "tooltip": "Smaller values use less VRAM, model default is 64, any other value will cause stutter"}),
"spatial_tile_sample_min_size": ("INT", {"default": 256, "min": 32, "max": 2048, "step": 32, "tooltip": "Spatial tile minimum size in pixels, smaller values use less VRAM, may introduce more seams"}),
"auto_tile_size": ("BOOLEAN", {"default": True, "tooltip": "Automatically set tile size based on defaults, above settings are ignored"}),
},
}
RETURN_TYPES = ("IMAGE",)
RETURN_NAMES = ("images",)
FUNCTION = "decode"
CATEGORY = "HunyuanVideoWrapper"
def decode(self, vae, samples, enable_vae_tiling, temporal_tiling_sample_size, spatial_tile_sample_min_size, auto_tile_size):
device = mm.get_torch_device()
offload_device = mm.unet_offload_device()
mm.soft_empty_cache()
latents = samples["samples"]
generator = torch.Generator(device=torch.device("cpu"))#.manual_seed(seed)
vae.to(device)
if not auto_tile_size:
vae.tile_latent_min_tsize = temporal_tiling_sample_size // 4
vae.tile_sample_min_size = spatial_tile_sample_min_size
vae.tile_latent_min_size = spatial_tile_sample_min_size // 8
if temporal_tiling_sample_size != 64:
vae.t_tile_overlap_factor = 0.0
else:
vae.t_tile_overlap_factor = 0.25
else:
#defaults
vae.tile_latent_min_tsize = 16
vae.tile_sample_min_size = 256
vae.tile_latent_min_size = 32
expand_temporal_dim = False
if len(latents.shape) == 4:
if isinstance(vae, AutoencoderKLCausal3D):
latents = latents.unsqueeze(2)
expand_temporal_dim = True
elif len(latents.shape) == 5:
pass
else:
raise ValueError(
f"Only support latents with shape (b, c, h, w) or (b, c, f, h, w), but got {latents.shape}."
)
latents = latents / vae.config.scaling_factor
latents = latents.to(vae.dtype).to(device)
if enable_vae_tiling:
vae.enable_tiling()
video = vae.decode(
latents, return_dict=False, generator=generator
)[0]
else:
video = vae.decode(
latents, return_dict=False, generator=generator
)[0]
if expand_temporal_dim or video.shape[2] == 1:
video = video.squeeze(2)
vae.to(offload_device)
mm.soft_empty_cache()
if len(video.shape) == 5:
video_processor = VideoProcessor(vae_scale_factor=8)
video_processor.config.do_resize = False
video = video_processor.postprocess_video(video=video, output_type="pt")
out = video[0].permute(0, 2, 3, 1).cpu().float()
else:
out = (video / 2 + 0.5).clamp(0, 1)
out = out.permute(0, 2, 3, 1).cpu().float()
return (out,)
#region VideoEncode
class HyVideoEncode:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"vae": ("VAE",),
"image": ("IMAGE",),
"enable_vae_tiling": ("BOOLEAN", {"default": True, "tooltip": "Drastically reduces memory use but may introduce seams"}),
"temporal_tiling_sample_size": ("INT", {"default": 16, "min": 4, "max": 256, "tooltip": "Smaller values use less VRAM, model default is 64, any other value will cause stutter"}),
"spatial_tile_sample_min_size": ("INT", {"default": 256, "min": 32, "max": 2048, "step": 32, "tooltip": "Spatial tile minimum size in pixels, smaller values use less VRAM, may introduce more seams"}),
"auto_tile_size": ("BOOLEAN", {"default": True, "tooltip": "Automatically set tile size based on defaults, above settings are ignored"}),
},
}
RETURN_TYPES = ("LATENT",)
RETURN_NAMES = ("samples",)
FUNCTION = "encode"
CATEGORY = "HunyuanVideoWrapper"