forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathconvert_ltx_to_diffusers.py
More file actions
371 lines (324 loc) 路 13.2 KB
/
Copy pathconvert_ltx_to_diffusers.py
File metadata and controls
371 lines (324 loc) 路 13.2 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
import argparse
from pathlib import Path
from typing import Any, Dict
import torch
from accelerate import init_empty_weights
from safetensors.torch import load_file
from transformers import T5EncoderModel, T5Tokenizer
from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
def remove_keys_(key: str, state_dict: Dict[str, Any]):
state_dict.pop(key)
TOKENIZER_MAX_LENGTH = 128
TRANSFORMER_KEYS_RENAME_DICT = {
"patchify_proj": "proj_in",
"adaln_single": "time_embed",
"q_norm": "norm_q",
"k_norm": "norm_k",
}
TRANSFORMER_SPECIAL_KEYS_REMAP = {
"vae": remove_keys_,
}
VAE_KEYS_RENAME_DICT = {
# decoder
"up_blocks.0": "mid_block",
"up_blocks.1": "up_blocks.0",
"up_blocks.2": "up_blocks.1.upsamplers.0",
"up_blocks.3": "up_blocks.1",
"up_blocks.4": "up_blocks.2.conv_in",
"up_blocks.5": "up_blocks.2.upsamplers.0",
"up_blocks.6": "up_blocks.2",
"up_blocks.7": "up_blocks.3.conv_in",
"up_blocks.8": "up_blocks.3.upsamplers.0",
"up_blocks.9": "up_blocks.3",
# encoder
"down_blocks.0": "down_blocks.0",
"down_blocks.1": "down_blocks.0.downsamplers.0",
"down_blocks.2": "down_blocks.0.conv_out",
"down_blocks.3": "down_blocks.1",
"down_blocks.4": "down_blocks.1.downsamplers.0",
"down_blocks.5": "down_blocks.1.conv_out",
"down_blocks.6": "down_blocks.2",
"down_blocks.7": "down_blocks.2.downsamplers.0",
"down_blocks.8": "down_blocks.3",
"down_blocks.9": "mid_block",
# common
"conv_shortcut": "conv_shortcut.conv",
"res_blocks": "resnets",
"norm3.norm": "norm3",
"per_channel_statistics.mean-of-means": "latents_mean",
"per_channel_statistics.std-of-means": "latents_std",
}
VAE_091_RENAME_DICT = {
# decoder
"up_blocks.0": "mid_block",
"up_blocks.1": "up_blocks.0.upsamplers.0",
"up_blocks.2": "up_blocks.0",
"up_blocks.3": "up_blocks.1.upsamplers.0",
"up_blocks.4": "up_blocks.1",
"up_blocks.5": "up_blocks.2.upsamplers.0",
"up_blocks.6": "up_blocks.2",
"up_blocks.7": "up_blocks.3.upsamplers.0",
"up_blocks.8": "up_blocks.3",
# common
"last_time_embedder": "time_embedder",
"last_scale_shift_table": "scale_shift_table",
}
VAE_095_RENAME_DICT = {
# decoder
"up_blocks.0": "mid_block",
"up_blocks.1": "up_blocks.0.upsamplers.0",
"up_blocks.2": "up_blocks.0",
"up_blocks.3": "up_blocks.1.upsamplers.0",
"up_blocks.4": "up_blocks.1",
"up_blocks.5": "up_blocks.2.upsamplers.0",
"up_blocks.6": "up_blocks.2",
"up_blocks.7": "up_blocks.3.upsamplers.0",
"up_blocks.8": "up_blocks.3",
# encoder
"down_blocks.0": "down_blocks.0",
"down_blocks.1": "down_blocks.0.downsamplers.0",
"down_blocks.2": "down_blocks.1",
"down_blocks.3": "down_blocks.1.downsamplers.0",
"down_blocks.4": "down_blocks.2",
"down_blocks.5": "down_blocks.2.downsamplers.0",
"down_blocks.6": "down_blocks.3",
"down_blocks.7": "down_blocks.3.downsamplers.0",
"down_blocks.8": "mid_block",
# common
"last_time_embedder": "time_embedder",
"last_scale_shift_table": "scale_shift_table",
}
VAE_SPECIAL_KEYS_REMAP = {
"per_channel_statistics.channel": remove_keys_,
"per_channel_statistics.mean-of-means": remove_keys_,
"per_channel_statistics.mean-of-stds": remove_keys_,
"model.diffusion_model": remove_keys_,
}
def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
state_dict = saved_dict
if "model" in saved_dict.keys():
state_dict = state_dict["model"]
if "module" in saved_dict.keys():
state_dict = state_dict["module"]
if "state_dict" in saved_dict.keys():
state_dict = state_dict["state_dict"]
return state_dict
def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
state_dict[new_key] = state_dict.pop(old_key)
def convert_transformer(
ckpt_path: str,
dtype: torch.dtype,
version: str = "0.9.0",
):
PREFIX_KEY = "model.diffusion_model."
original_state_dict = get_state_dict(load_file(ckpt_path))
config = {}
if version == "0.9.5":
config["_use_causal_rope_fix"] = True
with init_empty_weights():
transformer = LTXVideoTransformer3DModel(**config)
for key in list(original_state_dict.keys()):
new_key = key[:]
if new_key.startswith(PREFIX_KEY):
new_key = key[len(PREFIX_KEY) :]
for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
new_key = new_key.replace(replace_key, rename_key)
update_state_dict_inplace(original_state_dict, key, new_key)
for key in list(original_state_dict.keys()):
for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
if special_key not in key:
continue
handler_fn_inplace(key, original_state_dict)
transformer.load_state_dict(original_state_dict, strict=True, assign=True)
return transformer
def convert_vae(ckpt_path: str, config, dtype: torch.dtype):
PREFIX_KEY = "vae."
original_state_dict = get_state_dict(load_file(ckpt_path))
with init_empty_weights():
vae = AutoencoderKLLTXVideo(**config)
for key in list(original_state_dict.keys()):
new_key = key[:]
if new_key.startswith(PREFIX_KEY):
new_key = key[len(PREFIX_KEY) :]
for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
new_key = new_key.replace(replace_key, rename_key)
update_state_dict_inplace(original_state_dict, key, new_key)
for key in list(original_state_dict.keys()):
for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
if special_key not in key:
continue
handler_fn_inplace(key, original_state_dict)
vae.load_state_dict(original_state_dict, strict=True, assign=True)
return vae
def get_vae_config(version: str) -> Dict[str, Any]:
if version == "0.9.0":
config = {
"in_channels": 3,
"out_channels": 3,
"latent_channels": 128,
"block_out_channels": (128, 256, 512, 512),
"down_block_types": (
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
),
"decoder_block_out_channels": (128, 256, 512, 512),
"layers_per_block": (4, 3, 3, 3, 4),
"decoder_layers_per_block": (4, 3, 3, 3, 4),
"spatio_temporal_scaling": (True, True, True, False),
"decoder_spatio_temporal_scaling": (True, True, True, False),
"decoder_inject_noise": (False, False, False, False, False),
"downsample_type": ("conv", "conv", "conv", "conv"),
"upsample_residual": (False, False, False, False),
"upsample_factor": (1, 1, 1, 1),
"patch_size": 4,
"patch_size_t": 1,
"resnet_norm_eps": 1e-6,
"scaling_factor": 1.0,
"encoder_causal": True,
"decoder_causal": False,
"timestep_conditioning": False,
}
elif version == "0.9.1":
config = {
"in_channels": 3,
"out_channels": 3,
"latent_channels": 128,
"block_out_channels": (128, 256, 512, 512),
"down_block_types": (
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
"LTXVideoDownBlock3D",
),
"decoder_block_out_channels": (256, 512, 1024),
"layers_per_block": (4, 3, 3, 3, 4),
"decoder_layers_per_block": (5, 6, 7, 8),
"spatio_temporal_scaling": (True, True, True, False),
"decoder_spatio_temporal_scaling": (True, True, True),
"decoder_inject_noise": (True, True, True, False),
"downsample_type": ("conv", "conv", "conv", "conv"),
"upsample_residual": (True, True, True),
"upsample_factor": (2, 2, 2),
"timestep_conditioning": True,
"patch_size": 4,
"patch_size_t": 1,
"resnet_norm_eps": 1e-6,
"scaling_factor": 1.0,
"encoder_causal": True,
"decoder_causal": False,
}
VAE_KEYS_RENAME_DICT.update(VAE_091_RENAME_DICT)
elif version == "0.9.5":
config = {
"in_channels": 3,
"out_channels": 3,
"latent_channels": 128,
"block_out_channels": (128, 256, 512, 1024, 2048),
"down_block_types": (
"LTXVideo095DownBlock3D",
"LTXVideo095DownBlock3D",
"LTXVideo095DownBlock3D",
"LTXVideo095DownBlock3D",
),
"decoder_block_out_channels": (256, 512, 1024),
"layers_per_block": (4, 6, 6, 2, 2),
"decoder_layers_per_block": (5, 5, 5, 5),
"spatio_temporal_scaling": (True, True, True, True),
"decoder_spatio_temporal_scaling": (True, True, True),
"decoder_inject_noise": (False, False, False, False),
"downsample_type": ("spatial", "temporal", "spatiotemporal", "spatiotemporal"),
"upsample_residual": (True, True, True),
"upsample_factor": (2, 2, 2),
"timestep_conditioning": True,
"patch_size": 4,
"patch_size_t": 1,
"resnet_norm_eps": 1e-6,
"scaling_factor": 1.0,
"encoder_causal": True,
"decoder_causal": False,
"spatial_compression_ratio": 32,
"temporal_compression_ratio": 8,
}
VAE_KEYS_RENAME_DICT.update(VAE_095_RENAME_DICT)
return config
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
)
parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
parser.add_argument(
"--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
)
parser.add_argument(
"--typecast_text_encoder",
action="store_true",
default=False,
help="Whether or not to apply fp16/bf16 precision to text_encoder",
)
parser.add_argument("--save_pipeline", action="store_true")
parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
parser.add_argument(
"--version", type=str, default="0.9.0", choices=["0.9.0", "0.9.1", "0.9.5"], help="Version of the LTX model"
)
return parser.parse_args()
DTYPE_MAPPING = {
"fp32": torch.float32,
"fp16": torch.float16,
"bf16": torch.bfloat16,
}
VARIANT_MAPPING = {
"fp32": None,
"fp16": "fp16",
"bf16": "bf16",
}
if __name__ == "__main__":
args = get_args()
transformer = None
dtype = DTYPE_MAPPING[args.dtype]
variant = VARIANT_MAPPING[args.dtype]
output_path = Path(args.output_path)
if args.save_pipeline:
assert args.transformer_ckpt_path is not None and args.vae_ckpt_path is not None
if args.transformer_ckpt_path is not None:
transformer: LTXVideoTransformer3DModel = convert_transformer(args.transformer_ckpt_path, dtype)
if not args.save_pipeline:
transformer.save_pretrained(
output_path / "transformer", safe_serialization=True, max_shard_size="5GB", variant=variant
)
if args.vae_ckpt_path is not None:
config = get_vae_config(args.version)
vae: AutoencoderKLLTXVideo = convert_vae(args.vae_ckpt_path, config, dtype)
if not args.save_pipeline:
vae.save_pretrained(output_path / "vae", safe_serialization=True, max_shard_size="5GB", variant=variant)
if args.save_pipeline:
text_encoder_id = "google/t5-v1_1-xxl"
tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
if args.typecast_text_encoder:
text_encoder = text_encoder.to(dtype=dtype)
# Apparently, the conversion does not work anymore without this :shrug:
for param in text_encoder.parameters():
param.data = param.data.contiguous()
if args.version == "0.9.5":
scheduler = FlowMatchEulerDiscreteScheduler(use_dynamic_shifting=False)
else:
scheduler = FlowMatchEulerDiscreteScheduler(
use_dynamic_shifting=True,
base_shift=0.95,
max_shift=2.05,
base_image_seq_len=1024,
max_image_seq_len=4096,
shift_terminal=0.1,
)
pipe = LTXPipeline(
scheduler=scheduler,
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
transformer=transformer,
)
pipe.save_pretrained(args.output_path, safe_serialization=True, variant=variant, max_shard_size="5GB")