Spaces:
Runtime error
Runtime error
File size: 10,278 Bytes
77783a6 |
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 |
import torch
from huggingface_guess import model_list
from backend.diffusion_engine.base import ForgeDiffusionEngine, ForgeObjects
from backend.patcher.clip import CLIP
from backend.patcher.vae import VAE
from backend.patcher.unet import UnetPatcher
from backend.text_processing.classic_engine import ClassicTextProcessingEngine
from backend.args import dynamic_args
from backend import memory_management
from backend.nn.unet import Timestep
import safetensors.torch as sf
from backend import utils
from modules.shared import opts
class StableDiffusionXL(ForgeDiffusionEngine):
matched_guesses = [model_list.SDXL]
def __init__(self, estimated_config, huggingface_components):
super().__init__(estimated_config, huggingface_components)
clip = CLIP(
model_dict={
'clip_l': huggingface_components['text_encoder'],
'clip_g': huggingface_components['text_encoder_2']
},
tokenizer_dict={
'clip_l': huggingface_components['tokenizer'],
'clip_g': huggingface_components['tokenizer_2']
}
)
vae = VAE(model=huggingface_components['vae'])
unet = UnetPatcher.from_model(
model=huggingface_components['unet'],
diffusers_scheduler=huggingface_components['scheduler'],
config=estimated_config
)
self.text_processing_engine_l = ClassicTextProcessingEngine(
text_encoder=clip.cond_stage_model.clip_l,
tokenizer=clip.tokenizer.clip_l,
embedding_dir=dynamic_args['embedding_dir'],
embedding_key='clip_l',
embedding_expected_shape=2048,
emphasis_name=dynamic_args['emphasis_name'],
text_projection=False,
minimal_clip_skip=2,
clip_skip=2,
return_pooled=False,
final_layer_norm=False,
)
self.text_processing_engine_g = ClassicTextProcessingEngine(
text_encoder=clip.cond_stage_model.clip_g,
tokenizer=clip.tokenizer.clip_g,
embedding_dir=dynamic_args['embedding_dir'],
embedding_key='clip_g',
embedding_expected_shape=2048,
emphasis_name=dynamic_args['emphasis_name'],
text_projection=True,
minimal_clip_skip=2,
clip_skip=2,
return_pooled=True,
final_layer_norm=False,
)
self.embedder = Timestep(256)
self.forge_objects = ForgeObjects(unet=unet, clip=clip, vae=vae, clipvision=None)
self.forge_objects_original = self.forge_objects.shallow_copy()
self.forge_objects_after_applying_lora = self.forge_objects.shallow_copy()
# WebUI Legacy
self.is_sdxl = True
def set_clip_skip(self, clip_skip):
self.text_processing_engine_l.clip_skip = clip_skip
self.text_processing_engine_g.clip_skip = clip_skip
@torch.inference_mode()
def get_learned_conditioning(self, prompt: list[str]):
memory_management.load_model_gpu(self.forge_objects.clip.patcher)
cond_l = self.text_processing_engine_l(prompt)
cond_g, clip_pooled = self.text_processing_engine_g(prompt)
width = getattr(prompt, 'width', 1024) or 1024
height = getattr(prompt, 'height', 1024) or 1024
is_negative_prompt = getattr(prompt, 'is_negative_prompt', False)
crop_w = opts.sdxl_crop_left
crop_h = opts.sdxl_crop_top
target_width = width
target_height = height
out = [
self.embedder(torch.Tensor([height])), self.embedder(torch.Tensor([width])),
self.embedder(torch.Tensor([crop_h])), self.embedder(torch.Tensor([crop_w])),
self.embedder(torch.Tensor([target_height])), self.embedder(torch.Tensor([target_width]))
]
flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1).to(clip_pooled)
force_zero_negative_prompt = is_negative_prompt and all(x == '' for x in prompt)
if force_zero_negative_prompt:
clip_pooled = torch.zeros_like(clip_pooled)
cond_l = torch.zeros_like(cond_l)
cond_g = torch.zeros_like(cond_g)
cond = dict(
crossattn=torch.cat([cond_l, cond_g], dim=2),
vector=torch.cat([clip_pooled, flat], dim=1),
)
return cond
@torch.inference_mode()
def get_prompt_lengths_on_ui(self, prompt):
_, token_count = self.text_processing_engine_l.process_texts([prompt])
return token_count, self.text_processing_engine_l.get_target_prompt_token_count(token_count)
@torch.inference_mode()
def encode_first_stage(self, x):
sample = self.forge_objects.vae.encode(x.movedim(1, -1) * 0.5 + 0.5)
sample = self.forge_objects.vae.first_stage_model.process_in(sample)
return sample.to(x)
@torch.inference_mode()
def decode_first_stage(self, x):
sample = self.forge_objects.vae.first_stage_model.process_out(x)
sample = self.forge_objects.vae.decode(sample).movedim(-1, 1) * 2.0 - 1.0
return sample.to(x)
def save_checkpoint(self, filename):
sd = {}
sd.update(
utils.get_state_dict_after_quant(self.forge_objects.unet.model.diffusion_model, prefix='model.diffusion_model.')
)
sd.update(
model_list.SDXL.process_clip_state_dict_for_saving(self,
utils.get_state_dict_after_quant(self.forge_objects.clip.cond_stage_model, prefix='')
)
)
sd.update(
utils.get_state_dict_after_quant(self.forge_objects.vae.first_stage_model, prefix='first_stage_model.')
)
sf.save_file(sd, filename)
return filename
class StableDiffusionXLRefiner(ForgeDiffusionEngine):
matched_guesses = [model_list.SDXLRefiner]
def __init__(self, estimated_config, huggingface_components):
super().__init__(estimated_config, huggingface_components)
clip = CLIP(
model_dict={
'clip_g': huggingface_components['text_encoder']
},
tokenizer_dict={
'clip_g': huggingface_components['tokenizer'],
}
)
vae = VAE(model=huggingface_components['vae'])
unet = UnetPatcher.from_model(
model=huggingface_components['unet'],
diffusers_scheduler=huggingface_components['scheduler'],
config=estimated_config
)
self.text_processing_engine_g = ClassicTextProcessingEngine(
text_encoder=clip.cond_stage_model.clip_g,
tokenizer=clip.tokenizer.clip_g,
embedding_dir=dynamic_args['embedding_dir'],
embedding_key='clip_g',
embedding_expected_shape=2048,
emphasis_name=dynamic_args['emphasis_name'],
text_projection=True,
minimal_clip_skip=2,
clip_skip=2,
return_pooled=True,
final_layer_norm=False,
)
self.embedder = Timestep(256)
self.forge_objects = ForgeObjects(unet=unet, clip=clip, vae=vae, clipvision=None)
self.forge_objects_original = self.forge_objects.shallow_copy()
self.forge_objects_after_applying_lora = self.forge_objects.shallow_copy()
# WebUI Legacy
self.is_sdxl = True
def set_clip_skip(self, clip_skip):
self.text_processing_engine_g.clip_skip = clip_skip
@torch.inference_mode()
def get_learned_conditioning(self, prompt: list[str]):
memory_management.load_model_gpu(self.forge_objects.clip.patcher)
cond_g, clip_pooled = self.text_processing_engine_g(prompt)
width = getattr(prompt, 'width', 1024) or 1024
height = getattr(prompt, 'height', 1024) or 1024
is_negative_prompt = getattr(prompt, 'is_negative_prompt', False)
crop_w = opts.sdxl_crop_left
crop_h = opts.sdxl_crop_top
aesthetic = opts.sdxl_refiner_low_aesthetic_score if is_negative_prompt else opts.sdxl_refiner_high_aesthetic_score
out = [
self.embedder(torch.Tensor([height])), self.embedder(torch.Tensor([width])),
self.embedder(torch.Tensor([crop_h])), self.embedder(torch.Tensor([crop_w])),
self.embedder(torch.Tensor([aesthetic]))
]
flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1).to(clip_pooled)
force_zero_negative_prompt = is_negative_prompt and all(x == '' for x in prompt)
if force_zero_negative_prompt:
clip_pooled = torch.zeros_like(clip_pooled)
cond_g = torch.zeros_like(cond_g)
cond = dict(
crossattn=cond_g,
vector=torch.cat([clip_pooled, flat], dim=1),
)
return cond
@torch.inference_mode()
def get_prompt_lengths_on_ui(self, prompt):
_, token_count = self.text_processing_engine_g.process_texts([prompt])
return token_count, self.text_processing_engine_g.get_target_prompt_token_count(token_count)
@torch.inference_mode()
def encode_first_stage(self, x):
sample = self.forge_objects.vae.encode(x.movedim(1, -1) * 0.5 + 0.5)
sample = self.forge_objects.vae.first_stage_model.process_in(sample)
return sample.to(x)
@torch.inference_mode()
def decode_first_stage(self, x):
sample = self.forge_objects.vae.first_stage_model.process_out(x)
sample = self.forge_objects.vae.decode(sample).movedim(-1, 1) * 2.0 - 1.0
return sample.to(x)
def save_checkpoint(self, filename):
sd = {}
sd.update(
utils.get_state_dict_after_quant(self.forge_objects.unet.model.diffusion_model, prefix='model.diffusion_model.')
)
sd.update(
model_list.SDXLRefiner.process_clip_state_dict_for_saving(self,
utils.get_state_dict_after_quant(self.forge_objects.clip.cond_stage_model, prefix='')
)
)
sd.update(
utils.get_state_dict_after_quant(self.forge_objects.vae.first_stage_model, prefix='first_stage_model.')
)
sf.save_file(sd, filename)
return filename
|