{{ message }}
Repository navigation
Expand file tree
/
Copy pathconvert_cogview4_to_diffusers.py
More file actions
254 lines (212 loc) · 10.7 KB
/
Copy pathconvert_cogview4_to_diffusers.py
File metadata and controls
254 lines (212 loc) · 10.7 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
"""
Convert a CogView4 checkpoint from SAT(https://github.com/THUDM/SwissArmyTransformer) to the Diffusers format.
(deprecated Since 2025-02-07 and will remove it in later CogView4 version)
This script converts a CogView4 checkpoint to the Diffusers format, which can then be used
with the Diffusers library.
Example usage:
python scripts/convert_cogview4_to_diffusers.py \
--transformer_checkpoint_path 'your path/cogview4_6b/1/mp_rank_00_model_states.pt' \
--vae_checkpoint_path 'your path/cogview4_6b/imagekl_ch16.pt' \
--output_path "THUDM/CogView4-6B" \
--dtype "bf16"
Arguments:
--transformer_checkpoint_path: Path to Transformer state dict.
--vae_checkpoint_path: Path to VAE state dict.
--output_path: The path to save the converted model.
--push_to_hub: Whether to push the converted checkpoint to the HF Hub or not. Defaults to `False`.
--text_encoder_cache_dir: Cache directory where text encoder is located. Defaults to None, which means HF_HOME will be used
--dtype: The dtype to save the model in (default: "bf16", options: "fp16", "bf16", "fp32"). If None, the dtype of the state dict is considered.
Default is "bf16" because CogView4 uses bfloat16 for Training.
Note: You must provide either --original_state_dict_repo_id or --checkpoint_path.
"""
import argparse
from contextlib import nullcontext
import torch
from accelerate import init_empty_weights
from transformers import GlmForCausalLM, PreTrainedTokenizerFast
from diffusers import AutoencoderKL, CogView4Pipeline, CogView4Transformer2DModel, FlowMatchEulerDiscreteScheduler
from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
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("--transformer_checkpoint_path", default=None, type=str)
parser.add_argument("--vae_checkpoint_path", default=None, type=str)
parser.add_argument("--output_path", required=True, type=str)
parser.add_argument("--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving")
parser.add_argument("--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory")
parser.add_argument("--dtype", type=str, default="bf16")
args = parser.parse_args()
# this is specific to `AdaLayerNormContinuous`:
# diffusers implementation split the linear projection into the scale, shift while CogView4 split it tino shift, scale
def swap_scale_shift(weight, dim):
"""
Swap the scale and shift components in the weight tensor.
Args:
weight (torch.Tensor): The original weight tensor.
dim (int): The dimension along which to split.
Returns:
torch.Tensor: The modified weight tensor with scale and shift swapped.
"""
shift, scale = weight.chunk(2, dim=dim)
new_weight = torch.cat([scale, shift], dim=dim)
return new_weight
def convert_cogview4_transformer_checkpoint_to_diffusers(ckpt_path):
original_state_dict = torch.load(ckpt_path, map_location="cpu")
original_state_dict = original_state_dict["module"]
original_state_dict = {k.replace("model.diffusion_model.", ""): v for k, v in original_state_dict.items()}
new_state_dict = {}
# Convert patch_embed
new_state_dict["patch_embed.proj.weight"] = original_state_dict.pop("mixins.patch_embed.proj.weight")
new_state_dict["patch_embed.proj.bias"] = original_state_dict.pop("mixins.patch_embed.proj.bias")
new_state_dict["patch_embed.text_proj.weight"] = original_state_dict.pop("mixins.patch_embed.text_proj.weight")
new_state_dict["patch_embed.text_proj.bias"] = original_state_dict.pop("mixins.patch_embed.text_proj.bias")
# Convert time_condition_embed
new_state_dict["time_condition_embed.timestep_embedder.linear_1.weight"] = original_state_dict.pop(
"time_embed.0.weight"
)
new_state_dict["time_condition_embed.timestep_embedder.linear_1.bias"] = original_state_dict.pop(
"time_embed.0.bias"
)
new_state_dict["time_condition_embed.timestep_embedder.linear_2.weight"] = original_state_dict.pop(
"time_embed.2.weight"
)
new_state_dict["time_condition_embed.timestep_embedder.linear_2.bias"] = original_state_dict.pop(
"time_embed.2.bias"
)
new_state_dict["time_condition_embed.condition_embedder.linear_1.weight"] = original_state_dict.pop(
"label_emb.0.0.weight"
)
new_state_dict["time_condition_embed.condition_embedder.linear_1.bias"] = original_state_dict.pop(
"label_emb.0.0.bias"
)
new_state_dict["time_condition_embed.condition_embedder.linear_2.weight"] = original_state_dict.pop(
"label_emb.0.2.weight"
)
new_state_dict["time_condition_embed.condition_embedder.linear_2.bias"] = original_state_dict.pop(
"label_emb.0.2.bias"
)
# Convert transformer blocks, for cogview4 is 28 blocks
for i in range(28):
block_prefix = f"transformer_blocks.{i}."
old_prefix = f"transformer.layers.{i}."
adaln_prefix = f"mixins.adaln.adaln_modules.{i}."
new_state_dict[block_prefix + "norm1.linear.weight"] = original_state_dict.pop(adaln_prefix + "1.weight")
new_state_dict[block_prefix + "norm1.linear.bias"] = original_state_dict.pop(adaln_prefix + "1.bias")
qkv_weight = original_state_dict.pop(old_prefix + "attention.query_key_value.weight")
qkv_bias = original_state_dict.pop(old_prefix + "attention.query_key_value.bias")
q, k, v = qkv_weight.chunk(3, dim=0)
q_bias, k_bias, v_bias = qkv_bias.chunk(3, dim=0)
new_state_dict[block_prefix + "attn1.to_q.weight"] = q
new_state_dict[block_prefix + "attn1.to_q.bias"] = q_bias
new_state_dict[block_prefix + "attn1.to_k.weight"] = k
new_state_dict[block_prefix + "attn1.to_k.bias"] = k_bias
new_state_dict[block_prefix + "attn1.to_v.weight"] = v
new_state_dict[block_prefix + "attn1.to_v.bias"] = v_bias
new_state_dict[block_prefix + "attn1.to_out.0.weight"] = original_state_dict.pop(
old_prefix + "attention.dense.weight"
)
new_state_dict[block_prefix + "attn1.to_out.0.bias"] = original_state_dict.pop(
old_prefix + "attention.dense.bias"
)
new_state_dict[block_prefix + "ff.net.0.proj.weight"] = original_state_dict.pop(
old_prefix + "mlp.dense_h_to_4h.weight"
)
new_state_dict[block_prefix + "ff.net.0.proj.bias"] = original_state_dict.pop(
old_prefix + "mlp.dense_h_to_4h.bias"
)
new_state_dict[block_prefix + "ff.net.2.weight"] = original_state_dict.pop(
old_prefix + "mlp.dense_4h_to_h.weight"
)
new_state_dict[block_prefix + "ff.net.2.bias"] = original_state_dict.pop(old_prefix + "mlp.dense_4h_to_h.bias")
# Convert final norm and projection
new_state_dict["norm_out.linear.weight"] = swap_scale_shift(
original_state_dict.pop("mixins.final_layer.adaln.1.weight"), dim=0
)
new_state_dict["norm_out.linear.bias"] = swap_scale_shift(
original_state_dict.pop("mixins.final_layer.adaln.1.bias"), dim=0
)
new_state_dict["proj_out.weight"] = original_state_dict.pop("mixins.final_layer.linear.weight")
new_state_dict["proj_out.bias"] = original_state_dict.pop("mixins.final_layer.linear.bias")
return new_state_dict
def convert_cogview4_vae_checkpoint_to_diffusers(ckpt_path, vae_config):
original_state_dict = torch.load(ckpt_path, map_location="cpu")["state_dict"]
return convert_ldm_vae_checkpoint(original_state_dict, vae_config)
def main(args):
if 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}")
transformer = None
vae = None
if args.transformer_checkpoint_path is not None:
converted_transformer_state_dict = convert_cogview4_transformer_checkpoint_to_diffusers(
args.transformer_checkpoint_path
)
transformer = CogView4Transformer2DModel(
patch_size=2,
in_channels=16,
num_layers=28,
attention_head_dim=128,
num_attention_heads=32,
out_channels=16,
text_embed_dim=4096,
time_embed_dim=512,
condition_dim=256,
pos_embed_max_size=128,
)
transformer.load_state_dict(converted_transformer_state_dict, strict=True)
if dtype is not None:
# Original checkpoint data type will be preserved
transformer = transformer.to(dtype=dtype)
if args.vae_checkpoint_path is not None:
vae_config = {
"in_channels": 3,
"out_channels": 3,
"down_block_types": ("DownEncoderBlock2D",) * 4,
"up_block_types": ("UpDecoderBlock2D",) * 4,
"block_out_channels": (128, 512, 1024, 1024),
"layers_per_block": 3,
"act_fn": "silu",
"latent_channels": 16,
"norm_num_groups": 32,
"sample_size": 1024,
"scaling_factor": 1.0,
"shift_factor": 0.0,
"force_upcast": True,
"use_quant_conv": False,
"use_post_quant_conv": False,
"mid_block_add_attention": False,
}
converted_vae_state_dict = convert_cogview4_vae_checkpoint_to_diffusers(args.vae_checkpoint_path, vae_config)
vae = AutoencoderKL(**vae_config)
vae.load_state_dict(converted_vae_state_dict, strict=True)
if dtype is not None:
vae = vae.to(dtype=dtype)
text_encoder_id = "THUDM/glm-4-9b-hf"
tokenizer = PreTrainedTokenizerFast.from_pretrained(text_encoder_id)
text_encoder = GlmForCausalLM.from_pretrained(
text_encoder_id,
cache_dir=args.text_encoder_cache_dir,
torch_dtype=torch.bfloat16 if args.dtype == "bf16" else torch.float32,
)
for param in text_encoder.parameters():
param.data = param.data.contiguous()
scheduler = FlowMatchEulerDiscreteScheduler(
base_shift=0.25, max_shift=0.75, base_image_seq_len=256, use_dynamic_shifting=True, time_shift_type="linear"
)
pipe = CogView4Pipeline(
tokenizer=tokenizer,
text_encoder=text_encoder,
vae=vae,
transformer=transformer,
scheduler=scheduler,
)
# This is necessary for users with insufficient memory, such as those using Colab and notebooks, as it can
# save some memory used for model loading.
pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", push_to_hub=args.push_to_hub)
if __name__ == "__main__":
main(args)
You can’t perform that action at this time.