Upload 2 files
Browse files- convert_repo_to_safetensors_sd.py +127 -30
- convert_repo_to_safetensors_sd_gr.py +128 -30
convert_repo_to_safetensors_sd.py
CHANGED
@@ -1,12 +1,13 @@
|
|
1 |
# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
|
2 |
# *Only* converts the UNet, VAE, and Text Encoder.
|
3 |
# Does not convert optimizer state or any other thing.
|
4 |
-
# Written by jachiam
|
5 |
|
6 |
import argparse
|
7 |
import os.path as osp
|
|
|
8 |
|
9 |
import torch
|
|
|
10 |
|
11 |
|
12 |
# =================#
|
@@ -158,10 +159,21 @@ vae_conversion_map_attn = [
|
|
158 |
("proj_out.", "proj_attn."),
|
159 |
]
|
160 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
|
162 |
def reshape_weight_for_sd(w):
|
163 |
# convert HF linear weights to SD conv2d weights
|
164 |
-
|
|
|
|
|
|
|
165 |
|
166 |
|
167 |
def convert_vae_state_dict(vae_state_dict):
|
@@ -177,18 +189,92 @@ def convert_vae_state_dict(vae_state_dict):
|
|
177 |
mapping[k] = v
|
178 |
new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
|
179 |
weights_to_convert = ["q", "k", "v", "proj_out"]
|
|
|
180 |
for k, v in new_state_dict.items():
|
181 |
for weight_name in weights_to_convert:
|
182 |
if f"mid.attn_1.{weight_name}.weight" in k:
|
183 |
print(f"Reshaping {k} for SD format")
|
184 |
new_state_dict[k] = reshape_weight_for_sd(v)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
185 |
return new_state_dict
|
186 |
|
187 |
|
188 |
# =========================#
|
189 |
# Text Encoder Conversion #
|
190 |
# =========================#
|
191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
|
193 |
|
194 |
def convert_text_enc_state_dict(text_enc_dict):
|
@@ -196,45 +282,56 @@ def convert_text_enc_state_dict(text_enc_dict):
|
|
196 |
|
197 |
|
198 |
def convert_diffusers_to_safetensors(model_path, checkpoint_path, half = True):
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
213 |
|
214 |
# Convert the UNet model
|
215 |
-
unet_state_dict = torch.load(unet_path, map_location='cpu') if not input_safetensors else load_file(unet_path, device='cpu')
|
216 |
unet_state_dict = convert_unet_state_dict(unet_state_dict)
|
217 |
unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
|
218 |
|
219 |
# Convert the VAE model
|
220 |
-
vae_state_dict = torch.load(vae_path, map_location='cpu') if not input_safetensors else load_file(vae_path, device='cpu')
|
221 |
vae_state_dict = convert_vae_state_dict(vae_state_dict)
|
222 |
vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
|
223 |
|
224 |
-
#
|
225 |
-
|
226 |
-
|
227 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
228 |
|
229 |
# Put together new checkpoint
|
230 |
state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
|
231 |
if half:
|
232 |
-
state_dict = {k:v.half() for k,v in state_dict.items()}
|
233 |
-
|
234 |
-
|
235 |
-
else:
|
236 |
-
state_dict = {"state_dict": state_dict}
|
237 |
-
torch.save(state_dict, checkpoint_path)
|
238 |
|
239 |
|
240 |
def download_repo(repo_id, dir_path):
|
@@ -258,7 +355,7 @@ if __name__ == "__main__":
|
|
258 |
parser = argparse.ArgumentParser()
|
259 |
|
260 |
parser.add_argument("--repo_id", default=None, type=str, required=True, help="HF Repo ID of the model to convert.")
|
261 |
-
parser.add_argument("--half",
|
262 |
|
263 |
args = parser.parse_args()
|
264 |
assert args.repo_id is not None, "Must provide a Repo ID!"
|
|
|
1 |
# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
|
2 |
# *Only* converts the UNet, VAE, and Text Encoder.
|
3 |
# Does not convert optimizer state or any other thing.
|
|
|
4 |
|
5 |
import argparse
|
6 |
import os.path as osp
|
7 |
+
import re
|
8 |
|
9 |
import torch
|
10 |
+
from safetensors.torch import load_file, save_file
|
11 |
|
12 |
|
13 |
# =================#
|
|
|
159 |
("proj_out.", "proj_attn."),
|
160 |
]
|
161 |
|
162 |
+
# This is probably not the most ideal solution, but it does work.
|
163 |
+
vae_extra_conversion_map = [
|
164 |
+
("to_q", "q"),
|
165 |
+
("to_k", "k"),
|
166 |
+
("to_v", "v"),
|
167 |
+
("to_out.0", "proj_out"),
|
168 |
+
]
|
169 |
+
|
170 |
|
171 |
def reshape_weight_for_sd(w):
|
172 |
# convert HF linear weights to SD conv2d weights
|
173 |
+
if not w.ndim == 1:
|
174 |
+
return w.reshape(*w.shape, 1, 1)
|
175 |
+
else:
|
176 |
+
return w
|
177 |
|
178 |
|
179 |
def convert_vae_state_dict(vae_state_dict):
|
|
|
189 |
mapping[k] = v
|
190 |
new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
|
191 |
weights_to_convert = ["q", "k", "v", "proj_out"]
|
192 |
+
keys_to_rename = {}
|
193 |
for k, v in new_state_dict.items():
|
194 |
for weight_name in weights_to_convert:
|
195 |
if f"mid.attn_1.{weight_name}.weight" in k:
|
196 |
print(f"Reshaping {k} for SD format")
|
197 |
new_state_dict[k] = reshape_weight_for_sd(v)
|
198 |
+
for weight_name, real_weight_name in vae_extra_conversion_map:
|
199 |
+
if f"mid.attn_1.{weight_name}.weight" in k or f"mid.attn_1.{weight_name}.bias" in k:
|
200 |
+
keys_to_rename[k] = k.replace(weight_name, real_weight_name)
|
201 |
+
for k, v in keys_to_rename.items():
|
202 |
+
if k in new_state_dict:
|
203 |
+
print(f"Renaming {k} to {v}")
|
204 |
+
new_state_dict[v] = reshape_weight_for_sd(new_state_dict[k])
|
205 |
+
del new_state_dict[k]
|
206 |
return new_state_dict
|
207 |
|
208 |
|
209 |
# =========================#
|
210 |
# Text Encoder Conversion #
|
211 |
# =========================#
|
212 |
+
|
213 |
+
|
214 |
+
textenc_conversion_lst = [
|
215 |
+
# (stable-diffusion, HF Diffusers)
|
216 |
+
("resblocks.", "text_model.encoder.layers."),
|
217 |
+
("ln_1", "layer_norm1"),
|
218 |
+
("ln_2", "layer_norm2"),
|
219 |
+
(".c_fc.", ".fc1."),
|
220 |
+
(".c_proj.", ".fc2."),
|
221 |
+
(".attn", ".self_attn"),
|
222 |
+
("ln_final.", "transformer.text_model.final_layer_norm."),
|
223 |
+
("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
|
224 |
+
("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
|
225 |
+
]
|
226 |
+
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
|
227 |
+
textenc_pattern = re.compile("|".join(protected.keys()))
|
228 |
+
|
229 |
+
# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
|
230 |
+
code2idx = {"q": 0, "k": 1, "v": 2}
|
231 |
+
|
232 |
+
|
233 |
+
def convert_text_enc_state_dict_v20(text_enc_dict):
|
234 |
+
new_state_dict = {}
|
235 |
+
capture_qkv_weight = {}
|
236 |
+
capture_qkv_bias = {}
|
237 |
+
for k, v in text_enc_dict.items():
|
238 |
+
if (
|
239 |
+
k.endswith(".self_attn.q_proj.weight")
|
240 |
+
or k.endswith(".self_attn.k_proj.weight")
|
241 |
+
or k.endswith(".self_attn.v_proj.weight")
|
242 |
+
):
|
243 |
+
k_pre = k[: -len(".q_proj.weight")]
|
244 |
+
k_code = k[-len("q_proj.weight")]
|
245 |
+
if k_pre not in capture_qkv_weight:
|
246 |
+
capture_qkv_weight[k_pre] = [None, None, None]
|
247 |
+
capture_qkv_weight[k_pre][code2idx[k_code]] = v
|
248 |
+
continue
|
249 |
+
|
250 |
+
if (
|
251 |
+
k.endswith(".self_attn.q_proj.bias")
|
252 |
+
or k.endswith(".self_attn.k_proj.bias")
|
253 |
+
or k.endswith(".self_attn.v_proj.bias")
|
254 |
+
):
|
255 |
+
k_pre = k[: -len(".q_proj.bias")]
|
256 |
+
k_code = k[-len("q_proj.bias")]
|
257 |
+
if k_pre not in capture_qkv_bias:
|
258 |
+
capture_qkv_bias[k_pre] = [None, None, None]
|
259 |
+
capture_qkv_bias[k_pre][code2idx[k_code]] = v
|
260 |
+
continue
|
261 |
+
|
262 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
|
263 |
+
new_state_dict[relabelled_key] = v
|
264 |
+
|
265 |
+
for k_pre, tensors in capture_qkv_weight.items():
|
266 |
+
if None in tensors:
|
267 |
+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
|
268 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
|
269 |
+
new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)
|
270 |
+
|
271 |
+
for k_pre, tensors in capture_qkv_bias.items():
|
272 |
+
if None in tensors:
|
273 |
+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
|
274 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
|
275 |
+
new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)
|
276 |
+
|
277 |
+
return new_state_dict
|
278 |
|
279 |
|
280 |
def convert_text_enc_state_dict(text_enc_dict):
|
|
|
282 |
|
283 |
|
284 |
def convert_diffusers_to_safetensors(model_path, checkpoint_path, half = True):
|
285 |
+
# Path for safetensors
|
286 |
+
unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.safetensors")
|
287 |
+
vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.safetensors")
|
288 |
+
text_enc_path = osp.join(model_path, "text_encoder", "model.safetensors")
|
289 |
+
|
290 |
+
# Load models from safetensors if it exists, if it doesn't pytorch
|
291 |
+
if osp.exists(unet_path):
|
292 |
+
unet_state_dict = load_file(unet_path, device="cpu")
|
293 |
+
else:
|
294 |
+
unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.bin")
|
295 |
+
unet_state_dict = torch.load(unet_path, map_location="cpu")
|
296 |
+
|
297 |
+
if osp.exists(vae_path):
|
298 |
+
vae_state_dict = load_file(vae_path, device="cpu")
|
299 |
+
else:
|
300 |
+
vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.bin")
|
301 |
+
vae_state_dict = torch.load(vae_path, map_location="cpu")
|
302 |
+
|
303 |
+
if osp.exists(text_enc_path):
|
304 |
+
text_enc_dict = load_file(text_enc_path, device="cpu")
|
305 |
+
else:
|
306 |
+
text_enc_path = osp.join(model_path, "text_encoder", "pytorch_model.bin")
|
307 |
+
text_enc_dict = torch.load(text_enc_path, map_location="cpu")
|
308 |
|
309 |
# Convert the UNet model
|
|
|
310 |
unet_state_dict = convert_unet_state_dict(unet_state_dict)
|
311 |
unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
|
312 |
|
313 |
# Convert the VAE model
|
|
|
314 |
vae_state_dict = convert_vae_state_dict(vae_state_dict)
|
315 |
vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
|
316 |
|
317 |
+
# Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper
|
318 |
+
is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict
|
319 |
+
|
320 |
+
if is_v20_model:
|
321 |
+
# Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm
|
322 |
+
text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}
|
323 |
+
text_enc_dict = convert_text_enc_state_dict_v20(text_enc_dict)
|
324 |
+
text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}
|
325 |
+
else:
|
326 |
+
text_enc_dict = convert_text_enc_state_dict(text_enc_dict)
|
327 |
+
text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}
|
328 |
|
329 |
# Put together new checkpoint
|
330 |
state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
|
331 |
if half:
|
332 |
+
state_dict = {k: v.half() for k, v in state_dict.items()}
|
333 |
+
|
334 |
+
save_file(state_dict, checkpoint_path)
|
|
|
|
|
|
|
335 |
|
336 |
|
337 |
def download_repo(repo_id, dir_path):
|
|
|
355 |
parser = argparse.ArgumentParser()
|
356 |
|
357 |
parser.add_argument("--repo_id", default=None, type=str, required=True, help="HF Repo ID of the model to convert.")
|
358 |
+
parser.add_argument("--half", default=True, help="Save weights in half precision.")
|
359 |
|
360 |
args = parser.parse_args()
|
361 |
assert args.repo_id is not None, "Must provide a Repo ID!"
|
convert_repo_to_safetensors_sd_gr.py
CHANGED
@@ -1,14 +1,16 @@
|
|
1 |
# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
|
2 |
# *Only* converts the UNet, VAE, and Text Encoder.
|
3 |
# Does not convert optimizer state or any other thing.
|
4 |
-
# Written by jachiam
|
5 |
|
6 |
import argparse
|
7 |
import os.path as osp
|
|
|
8 |
|
9 |
import torch
|
|
|
10 |
import gradio as gr
|
11 |
|
|
|
12 |
# =================#
|
13 |
# UNet Conversion #
|
14 |
# =================#
|
@@ -158,10 +160,21 @@ vae_conversion_map_attn = [
|
|
158 |
("proj_out.", "proj_attn."),
|
159 |
]
|
160 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
161 |
|
162 |
def reshape_weight_for_sd(w):
|
163 |
# convert HF linear weights to SD conv2d weights
|
164 |
-
|
|
|
|
|
|
|
165 |
|
166 |
|
167 |
def convert_vae_state_dict(vae_state_dict):
|
@@ -177,18 +190,92 @@ def convert_vae_state_dict(vae_state_dict):
|
|
177 |
mapping[k] = v
|
178 |
new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
|
179 |
weights_to_convert = ["q", "k", "v", "proj_out"]
|
|
|
180 |
for k, v in new_state_dict.items():
|
181 |
for weight_name in weights_to_convert:
|
182 |
if f"mid.attn_1.{weight_name}.weight" in k:
|
183 |
print(f"Reshaping {k} for SD format")
|
184 |
new_state_dict[k] = reshape_weight_for_sd(v)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
185 |
return new_state_dict
|
186 |
|
187 |
|
188 |
# =========================#
|
189 |
# Text Encoder Conversion #
|
190 |
# =========================#
|
191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
|
193 |
|
194 |
def convert_text_enc_state_dict(text_enc_dict):
|
@@ -197,45 +284,56 @@ def convert_text_enc_state_dict(text_enc_dict):
|
|
197 |
|
198 |
def convert_diffusers_to_safetensors(model_path, checkpoint_path, half = True, progress=gr.Progress(track_tqdm=True)):
|
199 |
progress(0, desc="Start converting...")
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
214 |
|
215 |
# Convert the UNet model
|
216 |
-
unet_state_dict = torch.load(unet_path, map_location='cpu') if not input_safetensors else load_file(unet_path)
|
217 |
unet_state_dict = convert_unet_state_dict(unet_state_dict)
|
218 |
unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
|
219 |
|
220 |
# Convert the VAE model
|
221 |
-
vae_state_dict = torch.load(vae_path, map_location='cpu') if not input_safetensors else load_file(vae_path)
|
222 |
vae_state_dict = convert_vae_state_dict(vae_state_dict)
|
223 |
vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
|
224 |
|
225 |
-
#
|
226 |
-
|
227 |
-
|
228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
229 |
|
230 |
# Put together new checkpoint
|
231 |
state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
|
232 |
if half:
|
233 |
-
state_dict = {k:v.half() for k,v in state_dict.items()}
|
234 |
-
|
235 |
-
|
236 |
-
else:
|
237 |
-
state_dict = {"state_dict": state_dict}
|
238 |
-
torch.save(state_dict, checkpoint_path)
|
239 |
|
240 |
progress(1, desc="Converted.")
|
241 |
|
@@ -295,7 +393,7 @@ if __name__ == "__main__":
|
|
295 |
parser = argparse.ArgumentParser()
|
296 |
|
297 |
parser.add_argument("--repo_id", default=None, type=str, required=True, help="HF Repo ID of the model to convert.")
|
298 |
-
parser.add_argument("--half",
|
299 |
|
300 |
args = parser.parse_args()
|
301 |
assert args.repo_id is not None, "Must provide a Repo ID!"
|
|
|
1 |
# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
|
2 |
# *Only* converts the UNet, VAE, and Text Encoder.
|
3 |
# Does not convert optimizer state or any other thing.
|
|
|
4 |
|
5 |
import argparse
|
6 |
import os.path as osp
|
7 |
+
import re
|
8 |
|
9 |
import torch
|
10 |
+
from safetensors.torch import load_file, save_file
|
11 |
import gradio as gr
|
12 |
|
13 |
+
|
14 |
# =================#
|
15 |
# UNet Conversion #
|
16 |
# =================#
|
|
|
160 |
("proj_out.", "proj_attn."),
|
161 |
]
|
162 |
|
163 |
+
# This is probably not the most ideal solution, but it does work.
|
164 |
+
vae_extra_conversion_map = [
|
165 |
+
("to_q", "q"),
|
166 |
+
("to_k", "k"),
|
167 |
+
("to_v", "v"),
|
168 |
+
("to_out.0", "proj_out"),
|
169 |
+
]
|
170 |
+
|
171 |
|
172 |
def reshape_weight_for_sd(w):
|
173 |
# convert HF linear weights to SD conv2d weights
|
174 |
+
if not w.ndim == 1:
|
175 |
+
return w.reshape(*w.shape, 1, 1)
|
176 |
+
else:
|
177 |
+
return w
|
178 |
|
179 |
|
180 |
def convert_vae_state_dict(vae_state_dict):
|
|
|
190 |
mapping[k] = v
|
191 |
new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
|
192 |
weights_to_convert = ["q", "k", "v", "proj_out"]
|
193 |
+
keys_to_rename = {}
|
194 |
for k, v in new_state_dict.items():
|
195 |
for weight_name in weights_to_convert:
|
196 |
if f"mid.attn_1.{weight_name}.weight" in k:
|
197 |
print(f"Reshaping {k} for SD format")
|
198 |
new_state_dict[k] = reshape_weight_for_sd(v)
|
199 |
+
for weight_name, real_weight_name in vae_extra_conversion_map:
|
200 |
+
if f"mid.attn_1.{weight_name}.weight" in k or f"mid.attn_1.{weight_name}.bias" in k:
|
201 |
+
keys_to_rename[k] = k.replace(weight_name, real_weight_name)
|
202 |
+
for k, v in keys_to_rename.items():
|
203 |
+
if k in new_state_dict:
|
204 |
+
print(f"Renaming {k} to {v}")
|
205 |
+
new_state_dict[v] = reshape_weight_for_sd(new_state_dict[k])
|
206 |
+
del new_state_dict[k]
|
207 |
return new_state_dict
|
208 |
|
209 |
|
210 |
# =========================#
|
211 |
# Text Encoder Conversion #
|
212 |
# =========================#
|
213 |
+
|
214 |
+
|
215 |
+
textenc_conversion_lst = [
|
216 |
+
# (stable-diffusion, HF Diffusers)
|
217 |
+
("resblocks.", "text_model.encoder.layers."),
|
218 |
+
("ln_1", "layer_norm1"),
|
219 |
+
("ln_2", "layer_norm2"),
|
220 |
+
(".c_fc.", ".fc1."),
|
221 |
+
(".c_proj.", ".fc2."),
|
222 |
+
(".attn", ".self_attn"),
|
223 |
+
("ln_final.", "transformer.text_model.final_layer_norm."),
|
224 |
+
("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
|
225 |
+
("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
|
226 |
+
]
|
227 |
+
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
|
228 |
+
textenc_pattern = re.compile("|".join(protected.keys()))
|
229 |
+
|
230 |
+
# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
|
231 |
+
code2idx = {"q": 0, "k": 1, "v": 2}
|
232 |
+
|
233 |
+
|
234 |
+
def convert_text_enc_state_dict_v20(text_enc_dict):
|
235 |
+
new_state_dict = {}
|
236 |
+
capture_qkv_weight = {}
|
237 |
+
capture_qkv_bias = {}
|
238 |
+
for k, v in text_enc_dict.items():
|
239 |
+
if (
|
240 |
+
k.endswith(".self_attn.q_proj.weight")
|
241 |
+
or k.endswith(".self_attn.k_proj.weight")
|
242 |
+
or k.endswith(".self_attn.v_proj.weight")
|
243 |
+
):
|
244 |
+
k_pre = k[: -len(".q_proj.weight")]
|
245 |
+
k_code = k[-len("q_proj.weight")]
|
246 |
+
if k_pre not in capture_qkv_weight:
|
247 |
+
capture_qkv_weight[k_pre] = [None, None, None]
|
248 |
+
capture_qkv_weight[k_pre][code2idx[k_code]] = v
|
249 |
+
continue
|
250 |
+
|
251 |
+
if (
|
252 |
+
k.endswith(".self_attn.q_proj.bias")
|
253 |
+
or k.endswith(".self_attn.k_proj.bias")
|
254 |
+
or k.endswith(".self_attn.v_proj.bias")
|
255 |
+
):
|
256 |
+
k_pre = k[: -len(".q_proj.bias")]
|
257 |
+
k_code = k[-len("q_proj.bias")]
|
258 |
+
if k_pre not in capture_qkv_bias:
|
259 |
+
capture_qkv_bias[k_pre] = [None, None, None]
|
260 |
+
capture_qkv_bias[k_pre][code2idx[k_code]] = v
|
261 |
+
continue
|
262 |
+
|
263 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
|
264 |
+
new_state_dict[relabelled_key] = v
|
265 |
+
|
266 |
+
for k_pre, tensors in capture_qkv_weight.items():
|
267 |
+
if None in tensors:
|
268 |
+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
|
269 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
|
270 |
+
new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)
|
271 |
+
|
272 |
+
for k_pre, tensors in capture_qkv_bias.items():
|
273 |
+
if None in tensors:
|
274 |
+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
|
275 |
+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
|
276 |
+
new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)
|
277 |
+
|
278 |
+
return new_state_dict
|
279 |
|
280 |
|
281 |
def convert_text_enc_state_dict(text_enc_dict):
|
|
|
284 |
|
285 |
def convert_diffusers_to_safetensors(model_path, checkpoint_path, half = True, progress=gr.Progress(track_tqdm=True)):
|
286 |
progress(0, desc="Start converting...")
|
287 |
+
# Path for safetensors
|
288 |
+
unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.safetensors")
|
289 |
+
vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.safetensors")
|
290 |
+
text_enc_path = osp.join(model_path, "text_encoder", "model.safetensors")
|
291 |
+
|
292 |
+
# Load models from safetensors if it exists, if it doesn't pytorch
|
293 |
+
if osp.exists(unet_path):
|
294 |
+
unet_state_dict = load_file(unet_path, device="cpu")
|
295 |
+
else:
|
296 |
+
unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.bin")
|
297 |
+
unet_state_dict = torch.load(unet_path, map_location="cpu")
|
298 |
+
|
299 |
+
if osp.exists(vae_path):
|
300 |
+
vae_state_dict = load_file(vae_path, device="cpu")
|
301 |
+
else:
|
302 |
+
vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.bin")
|
303 |
+
vae_state_dict = torch.load(vae_path, map_location="cpu")
|
304 |
+
|
305 |
+
if osp.exists(text_enc_path):
|
306 |
+
text_enc_dict = load_file(text_enc_path, device="cpu")
|
307 |
+
else:
|
308 |
+
text_enc_path = osp.join(model_path, "text_encoder", "pytorch_model.bin")
|
309 |
+
text_enc_dict = torch.load(text_enc_path, map_location="cpu")
|
310 |
|
311 |
# Convert the UNet model
|
|
|
312 |
unet_state_dict = convert_unet_state_dict(unet_state_dict)
|
313 |
unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
|
314 |
|
315 |
# Convert the VAE model
|
|
|
316 |
vae_state_dict = convert_vae_state_dict(vae_state_dict)
|
317 |
vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
|
318 |
|
319 |
+
# Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper
|
320 |
+
is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict
|
321 |
+
|
322 |
+
if is_v20_model:
|
323 |
+
# Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm
|
324 |
+
text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}
|
325 |
+
text_enc_dict = convert_text_enc_state_dict_v20(text_enc_dict)
|
326 |
+
text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}
|
327 |
+
else:
|
328 |
+
text_enc_dict = convert_text_enc_state_dict(text_enc_dict)
|
329 |
+
text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}
|
330 |
|
331 |
# Put together new checkpoint
|
332 |
state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
|
333 |
if half:
|
334 |
+
state_dict = {k: v.half() for k, v in state_dict.items()}
|
335 |
+
|
336 |
+
save_file(state_dict, checkpoint_path)
|
|
|
|
|
|
|
337 |
|
338 |
progress(1, desc="Converted.")
|
339 |
|
|
|
393 |
parser = argparse.ArgumentParser()
|
394 |
|
395 |
parser.add_argument("--repo_id", default=None, type=str, required=True, help="HF Repo ID of the model to convert.")
|
396 |
+
parser.add_argument("--half", default=True, help="Save weights in half precision.")
|
397 |
|
398 |
args = parser.parse_args()
|
399 |
assert args.repo_id is not None, "Must provide a Repo ID!"
|