Repository navigation
Expand file tree
/
Copy pathconvert_sd3_to_diffusers.py
More file actions
351 lines (297 loc) 路 15.9 KB
/
Copy pathconvert_sd3_to_diffusers.py
File metadata and controls
351 lines (297 loc) 路 15.9 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
import argparse
from contextlib import nullcontext
import safetensors.torch
import torch
from accelerate import init_empty_weights
from diffusers import AutoencoderKL, SD3Transformer2DModel
from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
from diffusers.models.model_loading_utils import load_model_dict_into_meta
from diffusers.utils.import_utils import is_accelerate_available
CTX = init_empty_weights if is_accelerate_available() else nullcontext
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint_path", type=str)
parser.add_argument("--output_path", type=str)
parser.add_argument("--dtype", type=str)
args = parser.parse_args()
def load_original_checkpoint(ckpt_path):
original_state_dict = safetensors.torch.load_file(ckpt_path)
keys = list(original_state_dict.keys())
for k in keys:
if "model.diffusion_model." in k:
original_state_dict[k.replace("model.diffusion_model.", "")] = original_state_dict.pop(k)
return original_state_dict
# in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
# while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
def swap_scale_shift(weight, dim):
shift, scale = weight.chunk(2, dim=0)
new_weight = torch.cat([scale, shift], dim=0)
return new_weight
def convert_sd3_transformer_checkpoint_to_diffusers(
original_state_dict, num_layers, caption_projection_dim, dual_attention_layers, has_qk_norm
):
converted_state_dict = {}
# Positional and patch embeddings.
converted_state_dict["pos_embed.pos_embed"] = original_state_dict.pop("pos_embed")
converted_state_dict["pos_embed.proj.weight"] = original_state_dict.pop("x_embedder.proj.weight")
converted_state_dict["pos_embed.proj.bias"] = original_state_dict.pop("x_embedder.proj.bias")
# Timestep embeddings.
converted_state_dict["time_text_embed.timestep_embedder.linear_1.weight"] = original_state_dict.pop(
"t_embedder.mlp.0.weight"
)
converted_state_dict["time_text_embed.timestep_embedder.linear_1.bias"] = original_state_dict.pop(
"t_embedder.mlp.0.bias"
)
converted_state_dict["time_text_embed.timestep_embedder.linear_2.weight"] = original_state_dict.pop(
"t_embedder.mlp.2.weight"
)
converted_state_dict["time_text_embed.timestep_embedder.linear_2.bias"] = original_state_dict.pop(
"t_embedder.mlp.2.bias"
)
# Context projections.
converted_state_dict["context_embedder.weight"] = original_state_dict.pop("context_embedder.weight")
converted_state_dict["context_embedder.bias"] = original_state_dict.pop("context_embedder.bias")
# Pooled context projection.
converted_state_dict["time_text_embed.text_embedder.linear_1.weight"] = original_state_dict.pop(
"y_embedder.mlp.0.weight"
)
converted_state_dict["time_text_embed.text_embedder.linear_1.bias"] = original_state_dict.pop(
"y_embedder.mlp.0.bias"
)
converted_state_dict["time_text_embed.text_embedder.linear_2.weight"] = original_state_dict.pop(
"y_embedder.mlp.2.weight"
)
converted_state_dict["time_text_embed.text_embedder.linear_2.bias"] = original_state_dict.pop(
"y_embedder.mlp.2.bias"
)
# Transformer blocks 馃幐.
for i in range(num_layers):
# Q, K, V
sample_q, sample_k, sample_v = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.x_block.attn.qkv.weight"), 3, dim=0
)
context_q, context_k, context_v = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.context_block.attn.qkv.weight"), 3, dim=0
)
sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.x_block.attn.qkv.bias"), 3, dim=0
)
context_q_bias, context_k_bias, context_v_bias = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.context_block.attn.qkv.bias"), 3, dim=0
)
converted_state_dict[f"transformer_blocks.{i}.attn.to_q.weight"] = torch.cat([sample_q])
converted_state_dict[f"transformer_blocks.{i}.attn.to_q.bias"] = torch.cat([sample_q_bias])
converted_state_dict[f"transformer_blocks.{i}.attn.to_k.weight"] = torch.cat([sample_k])
converted_state_dict[f"transformer_blocks.{i}.attn.to_k.bias"] = torch.cat([sample_k_bias])
converted_state_dict[f"transformer_blocks.{i}.attn.to_v.weight"] = torch.cat([sample_v])
converted_state_dict[f"transformer_blocks.{i}.attn.to_v.bias"] = torch.cat([sample_v_bias])
converted_state_dict[f"transformer_blocks.{i}.attn.add_q_proj.weight"] = torch.cat([context_q])
converted_state_dict[f"transformer_blocks.{i}.attn.add_q_proj.bias"] = torch.cat([context_q_bias])
converted_state_dict[f"transformer_blocks.{i}.attn.add_k_proj.weight"] = torch.cat([context_k])
converted_state_dict[f"transformer_blocks.{i}.attn.add_k_proj.bias"] = torch.cat([context_k_bias])
converted_state_dict[f"transformer_blocks.{i}.attn.add_v_proj.weight"] = torch.cat([context_v])
converted_state_dict[f"transformer_blocks.{i}.attn.add_v_proj.bias"] = torch.cat([context_v_bias])
# qk norm
if has_qk_norm:
converted_state_dict[f"transformer_blocks.{i}.attn.norm_q.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn.ln_q.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn.norm_k.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn.ln_k.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn.norm_added_q.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.attn.ln_q.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn.norm_added_k.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.attn.ln_k.weight"
)
# output projections.
converted_state_dict[f"transformer_blocks.{i}.attn.to_out.0.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn.proj.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn.to_out.0.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn.proj.bias"
)
if not (i == num_layers - 1):
converted_state_dict[f"transformer_blocks.{i}.attn.to_add_out.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.attn.proj.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn.to_add_out.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.attn.proj.bias"
)
# attn2
if i in dual_attention_layers:
# Q, K, V
sample_q2, sample_k2, sample_v2 = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.x_block.attn2.qkv.weight"), 3, dim=0
)
sample_q2_bias, sample_k2_bias, sample_v2_bias = torch.chunk(
original_state_dict.pop(f"joint_blocks.{i}.x_block.attn2.qkv.bias"), 3, dim=0
)
converted_state_dict[f"transformer_blocks.{i}.attn2.to_q.weight"] = torch.cat([sample_q2])
converted_state_dict[f"transformer_blocks.{i}.attn2.to_q.bias"] = torch.cat([sample_q2_bias])
converted_state_dict[f"transformer_blocks.{i}.attn2.to_k.weight"] = torch.cat([sample_k2])
converted_state_dict[f"transformer_blocks.{i}.attn2.to_k.bias"] = torch.cat([sample_k2_bias])
converted_state_dict[f"transformer_blocks.{i}.attn2.to_v.weight"] = torch.cat([sample_v2])
converted_state_dict[f"transformer_blocks.{i}.attn2.to_v.bias"] = torch.cat([sample_v2_bias])
# qk norm
if has_qk_norm:
converted_state_dict[f"transformer_blocks.{i}.attn2.norm_q.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn2.ln_q.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn2.norm_k.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn2.ln_k.weight"
)
# output projections.
converted_state_dict[f"transformer_blocks.{i}.attn2.to_out.0.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn2.proj.weight"
)
converted_state_dict[f"transformer_blocks.{i}.attn2.to_out.0.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.attn2.proj.bias"
)
# norms.
converted_state_dict[f"transformer_blocks.{i}.norm1.linear.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.adaLN_modulation.1.weight"
)
converted_state_dict[f"transformer_blocks.{i}.norm1.linear.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.adaLN_modulation.1.bias"
)
if not (i == num_layers - 1):
converted_state_dict[f"transformer_blocks.{i}.norm1_context.linear.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.adaLN_modulation.1.weight"
)
converted_state_dict[f"transformer_blocks.{i}.norm1_context.linear.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.adaLN_modulation.1.bias"
)
else:
converted_state_dict[f"transformer_blocks.{i}.norm1_context.linear.weight"] = swap_scale_shift(
original_state_dict.pop(f"joint_blocks.{i}.context_block.adaLN_modulation.1.weight"),
dim=caption_projection_dim,
)
converted_state_dict[f"transformer_blocks.{i}.norm1_context.linear.bias"] = swap_scale_shift(
original_state_dict.pop(f"joint_blocks.{i}.context_block.adaLN_modulation.1.bias"),
dim=caption_projection_dim,
)
# ffs.
converted_state_dict[f"transformer_blocks.{i}.ff.net.0.proj.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.mlp.fc1.weight"
)
converted_state_dict[f"transformer_blocks.{i}.ff.net.0.proj.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.mlp.fc1.bias"
)
converted_state_dict[f"transformer_blocks.{i}.ff.net.2.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.mlp.fc2.weight"
)
converted_state_dict[f"transformer_blocks.{i}.ff.net.2.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.x_block.mlp.fc2.bias"
)
if not (i == num_layers - 1):
converted_state_dict[f"transformer_blocks.{i}.ff_context.net.0.proj.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.mlp.fc1.weight"
)
converted_state_dict[f"transformer_blocks.{i}.ff_context.net.0.proj.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.mlp.fc1.bias"
)
converted_state_dict[f"transformer_blocks.{i}.ff_context.net.2.weight"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.mlp.fc2.weight"
)
converted_state_dict[f"transformer_blocks.{i}.ff_context.net.2.bias"] = original_state_dict.pop(
f"joint_blocks.{i}.context_block.mlp.fc2.bias"
)
# Final blocks.
converted_state_dict["proj_out.weight"] = original_state_dict.pop("final_layer.linear.weight")
converted_state_dict["proj_out.bias"] = original_state_dict.pop("final_layer.linear.bias")
converted_state_dict["norm_out.linear.weight"] = swap_scale_shift(
original_state_dict.pop("final_layer.adaLN_modulation.1.weight"), dim=caption_projection_dim
)
converted_state_dict["norm_out.linear.bias"] = swap_scale_shift(
original_state_dict.pop("final_layer.adaLN_modulation.1.bias"), dim=caption_projection_dim
)
return converted_state_dict
def is_vae_in_checkpoint(original_state_dict):
return ("first_stage_model.decoder.conv_in.weight" in original_state_dict) and (
"first_stage_model.encoder.conv_in.weight" in original_state_dict
)
def get_attn2_layers(state_dict):
attn2_layers = []
for key in state_dict.keys():
if "attn2." in key:
# Extract the layer number from the key
layer_num = int(key.split(".")[1])
attn2_layers.append(layer_num)
return tuple(sorted(set(attn2_layers)))
def get_pos_embed_max_size(state_dict):
num_patches = state_dict["pos_embed"].shape[1]
pos_embed_max_size = int(num_patches**0.5)
return pos_embed_max_size
def get_caption_projection_dim(state_dict):
caption_projection_dim = state_dict["context_embedder.weight"].shape[0]
return caption_projection_dim
def main(args):
original_ckpt = load_original_checkpoint(args.checkpoint_path)
original_dtype = next(iter(original_ckpt.values())).dtype
# Initialize dtype with a default value
dtype = None
if args.dtype is None:
dtype = original_dtype
elif args.dtype == "fp16":
dtype = torch.float16
elif args.dtype == "bf16":
dtype = torch.bfloat16
elif args.dtype == "fp32":
dtype = torch.float32
else:
raise ValueError(f"Unsupported dtype: {args.dtype}")
if dtype != original_dtype:
print(
f"Checkpoint dtype {original_dtype} does not match requested dtype {dtype}. This can lead to unexpected results, proceed with caution."
)
num_layers = list(set(int(k.split(".", 2)[1]) for k in original_ckpt if "joint_blocks" in k))[-1] + 1 # noqa: C401
caption_projection_dim = get_caption_projection_dim(original_ckpt)
# () for sd3.0; (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12) for sd3.5
attn2_layers = get_attn2_layers(original_ckpt)
# sd3.5 use qk norm("rms_norm")
has_qk_norm = any("ln_q" in key for key in original_ckpt.keys())
# sd3.5 2b use pox_embed_max_size=384 and sd3.0 and sd3.5 8b use 192
pos_embed_max_size = get_pos_embed_max_size(original_ckpt)
converted_transformer_state_dict = convert_sd3_transformer_checkpoint_to_diffusers(
original_ckpt, num_layers, caption_projection_dim, attn2_layers, has_qk_norm
)
with CTX():
transformer = SD3Transformer2DModel(
sample_size=128,
patch_size=2,
in_channels=16,
joint_attention_dim=4096,
num_layers=num_layers,
caption_projection_dim=caption_projection_dim,
num_attention_heads=num_layers,
pos_embed_max_size=pos_embed_max_size,
qk_norm="rms_norm" if has_qk_norm else None,
dual_attention_layers=attn2_layers,
)
if is_accelerate_available():
load_model_dict_into_meta(transformer, converted_transformer_state_dict)
else:
transformer.load_state_dict(converted_transformer_state_dict, strict=True)
print("Saving SD3 Transformer in Diffusers format.")
transformer.to(dtype).save_pretrained(f"{args.output_path}/transformer")
if is_vae_in_checkpoint(original_ckpt):
with CTX():
vae = AutoencoderKL.from_config(
"stabilityai/stable-diffusion-xl-base-1.0",
subfolder="vae",
latent_channels=16,
use_post_quant_conv=False,
use_quant_conv=False,
scaling_factor=1.5305,
shift_factor=0.0609,
)
converted_vae_state_dict = convert_ldm_vae_checkpoint(original_ckpt, vae.config)
if is_accelerate_available():
load_model_dict_into_meta(vae, converted_vae_state_dict)
else:
vae.load_state_dict(converted_vae_state_dict, strict=True)
print("Saving SD3 Autoencoder in Diffusers format.")
vae.to(dtype).save_pretrained(f"{args.output_path}/vae")
if __name__ == "__main__":
main(args)