{{ message }}
Repository navigation
Expand file tree
/
Copy pathconvert_ovis_image_to_diffusers.py
More file actions
263 lines (233 loc) · 12.4 KB
/
Copy pathconvert_ovis_image_to_diffusers.py
File metadata and controls
263 lines (233 loc) · 12.4 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
import argparse
from contextlib import nullcontext
import safetensors.torch
import torch
from accelerate import init_empty_weights
from huggingface_hub import hf_hub_download
from diffusers import OvisImageTransformer2DModel
from diffusers.utils.import_utils import is_accelerate_available
"""
# Transformer
python scripts/convert_ovis_image_to_diffusers.py \
--original_state_dict_repo_id "AIDC-AI/Ovis-Image-7B" \
--filename "ovis_image.safetensors"
--output_path "ovis-image" \
--transformer
"""
CTX = init_empty_weights if is_accelerate_available() else nullcontext
parser = argparse.ArgumentParser()
parser.add_argument("--original_state_dict_repo_id", default=None, type=str)
parser.add_argument("--filename", default="ovis_image.safetensors", type=str)
parser.add_argument("--checkpoint_path", default=None, type=str)
parser.add_argument("--in_channels", type=int, default=64)
parser.add_argument("--out_channels", type=int, default=None)
parser.add_argument("--transformer", action="store_true")
parser.add_argument("--output_path", type=str)
parser.add_argument("--dtype", type=str, default="bf16")
args = parser.parse_args()
dtype = torch.bfloat16 if args.dtype == "bf16" else torch.float32
def load_original_checkpoint(args):
if args.original_state_dict_repo_id is not None:
ckpt_path = hf_hub_download(repo_id=args.original_state_dict_repo_id, filename=args.filename)
elif args.checkpoint_path is not None:
ckpt_path = args.checkpoint_path
else:
raise ValueError(" please provide either `original_state_dict_repo_id` or a local `checkpoint_path`")
original_state_dict = safetensors.torch.load_file(ckpt_path)
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):
shift, scale = weight.chunk(2, dim=0)
new_weight = torch.cat([scale, shift], dim=0)
return new_weight
def convert_ovis_image_transformer_checkpoint_to_diffusers(
original_state_dict, num_layers, num_single_layers, inner_dim, mlp_ratio=4.0
):
converted_state_dict = {}
## time_text_embed.timestep_embedder <- time_in
converted_state_dict["timestep_embedder.linear_1.weight"] = original_state_dict.pop("time_in.in_layer.weight")
converted_state_dict["timestep_embedder.linear_1.bias"] = original_state_dict.pop("time_in.in_layer.bias")
converted_state_dict["timestep_embedder.linear_2.weight"] = original_state_dict.pop("time_in.out_layer.weight")
converted_state_dict["timestep_embedder.linear_2.bias"] = original_state_dict.pop("time_in.out_layer.bias")
# context_embedder
converted_state_dict["context_embedder_norm.weight"] = original_state_dict.pop("semantic_txt_norm.weight")
converted_state_dict["context_embedder.weight"] = original_state_dict.pop("semantic_txt_in.weight")
converted_state_dict["context_embedder.bias"] = original_state_dict.pop("semantic_txt_in.bias")
# x_embedder
converted_state_dict["x_embedder.weight"] = original_state_dict.pop("img_in.weight")
converted_state_dict["x_embedder.bias"] = original_state_dict.pop("img_in.bias")
# double transformer blocks
for i in range(num_layers):
block_prefix = f"transformer_blocks.{i}."
# norms.
## norm1
converted_state_dict[f"{block_prefix}norm1.linear.weight"] = original_state_dict.pop(
f"double_blocks.{i}.img_mod.lin.weight"
)
converted_state_dict[f"{block_prefix}norm1.linear.bias"] = original_state_dict.pop(
f"double_blocks.{i}.img_mod.lin.bias"
)
## norm1_context
converted_state_dict[f"{block_prefix}norm1_context.linear.weight"] = original_state_dict.pop(
f"double_blocks.{i}.txt_mod.lin.weight"
)
converted_state_dict[f"{block_prefix}norm1_context.linear.bias"] = original_state_dict.pop(
f"double_blocks.{i}.txt_mod.lin.bias"
)
# Q, K, V
sample_q, sample_k, sample_v = torch.chunk(
original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0
)
context_q, context_k, context_v = torch.chunk(
original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0
)
sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0
)
context_q_bias, context_k_bias, context_v_bias = torch.chunk(
original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0
)
converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q])
converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias])
converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k])
converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias])
converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v])
converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias])
converted_state_dict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q])
converted_state_dict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias])
converted_state_dict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k])
converted_state_dict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias])
converted_state_dict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v])
converted_state_dict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias])
# qk_norm
converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop(
f"double_blocks.{i}.img_attn.norm.query_norm.weight"
)
converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop(
f"double_blocks.{i}.img_attn.norm.key_norm.weight"
)
converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = original_state_dict.pop(
f"double_blocks.{i}.txt_attn.norm.query_norm.weight"
)
converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = original_state_dict.pop(
f"double_blocks.{i}.txt_attn.norm.key_norm.weight"
)
# ff img_mlp
converted_state_dict[f"{block_prefix}ff.net.0.proj.weight"] = torch.cat(
[
original_state_dict.pop(f"double_blocks.{i}.img_mlp.up_proj.weight"),
original_state_dict.pop(f"double_blocks.{i}.img_mlp.gate_proj.weight"),
],
dim=0,
)
converted_state_dict[f"{block_prefix}ff.net.0.proj.bias"] = torch.cat(
[
original_state_dict.pop(f"double_blocks.{i}.img_mlp.up_proj.bias"),
original_state_dict.pop(f"double_blocks.{i}.img_mlp.gate_proj.bias"),
],
dim=0,
)
converted_state_dict[f"{block_prefix}ff.net.2.weight"] = original_state_dict.pop(
f"double_blocks.{i}.img_mlp.down_proj.weight"
)
converted_state_dict[f"{block_prefix}ff.net.2.bias"] = original_state_dict.pop(
f"double_blocks.{i}.img_mlp.down_proj.bias"
)
converted_state_dict[f"{block_prefix}ff_context.net.0.proj.weight"] = torch.cat(
[
original_state_dict.pop(f"double_blocks.{i}.txt_mlp.up_proj.weight"),
original_state_dict.pop(f"double_blocks.{i}.txt_mlp.gate_proj.weight"),
],
dim=0,
)
converted_state_dict[f"{block_prefix}ff_context.net.0.proj.bias"] = torch.cat(
[
original_state_dict.pop(f"double_blocks.{i}.txt_mlp.up_proj.bias"),
original_state_dict.pop(f"double_blocks.{i}.txt_mlp.gate_proj.bias"),
],
dim=0,
)
converted_state_dict[f"{block_prefix}ff_context.net.2.weight"] = original_state_dict.pop(
f"double_blocks.{i}.txt_mlp.down_proj.weight"
)
converted_state_dict[f"{block_prefix}ff_context.net.2.bias"] = original_state_dict.pop(
f"double_blocks.{i}.txt_mlp.down_proj.bias"
)
# output projections.
converted_state_dict[f"{block_prefix}attn.to_out.0.weight"] = original_state_dict.pop(
f"double_blocks.{i}.img_attn.proj.weight"
)
converted_state_dict[f"{block_prefix}attn.to_out.0.bias"] = original_state_dict.pop(
f"double_blocks.{i}.img_attn.proj.bias"
)
converted_state_dict[f"{block_prefix}attn.to_add_out.weight"] = original_state_dict.pop(
f"double_blocks.{i}.txt_attn.proj.weight"
)
converted_state_dict[f"{block_prefix}attn.to_add_out.bias"] = original_state_dict.pop(
f"double_blocks.{i}.txt_attn.proj.bias"
)
# single transformer blocks
for i in range(num_single_layers):
block_prefix = f"single_transformer_blocks.{i}."
# norm.linear <- single_blocks.0.modulation.lin
converted_state_dict[f"{block_prefix}norm.linear.weight"] = original_state_dict.pop(
f"single_blocks.{i}.modulation.lin.weight"
)
converted_state_dict[f"{block_prefix}norm.linear.bias"] = original_state_dict.pop(
f"single_blocks.{i}.modulation.lin.bias"
)
# Q, K, V, mlp
mlp_hidden_dim = int(inner_dim * mlp_ratio)
split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim * 2)
q, k, v, mlp = torch.split(original_state_dict.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0)
q_bias, k_bias, v_bias, mlp_bias = torch.split(
original_state_dict.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0
)
converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q])
converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias])
converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k])
converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias])
converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v])
converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias])
converted_state_dict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp])
converted_state_dict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias])
# qk norm
converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop(
f"single_blocks.{i}.norm.query_norm.weight"
)
converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop(
f"single_blocks.{i}.norm.key_norm.weight"
)
# output projections.
converted_state_dict[f"{block_prefix}proj_out.weight"] = original_state_dict.pop(
f"single_blocks.{i}.linear2.weight"
)
converted_state_dict[f"{block_prefix}proj_out.bias"] = original_state_dict.pop(
f"single_blocks.{i}.linear2.bias"
)
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")
)
converted_state_dict["norm_out.linear.bias"] = swap_scale_shift(
original_state_dict.pop("final_layer.adaLN_modulation.1.bias")
)
return converted_state_dict
def main(args):
original_ckpt = load_original_checkpoint(args)
if args.transformer:
num_layers = 6
num_single_layers = 27
inner_dim = 3072
mlp_ratio = 4.0
converted_transformer_state_dict = convert_ovis_image_transformer_checkpoint_to_diffusers(
original_ckpt, num_layers, num_single_layers, inner_dim, mlp_ratio=mlp_ratio
)
transformer = OvisImageTransformer2DModel(in_channels=args.in_channels, out_channels=args.out_channels)
transformer.load_state_dict(converted_transformer_state_dict, strict=True)
print("Saving Ovis-Image Transformer in Diffusers format.")
transformer.to(dtype).save_pretrained(f"{args.output_path}/transformer")
if __name__ == "__main__":
main(args)
You can’t perform that action at this time.