Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -14,7 +14,8 @@ from tqdm import tqdm
|
|
14 |
import bitsandbytes as bnb
|
15 |
from bitsandbytes.nn.modules import Params4bit, QuantState
|
16 |
from transformers import (
|
17 |
-
|
|
|
18 |
CLIPTextModel, CLIPTokenizer,
|
19 |
T5EncoderModel, T5Tokenizer
|
20 |
)
|
@@ -23,17 +24,27 @@ from huggingface_hub import hf_hub_download
|
|
23 |
from safetensors.torch import load_file
|
24 |
from einops import rearrange, repeat
|
25 |
|
26 |
-
# 1) ์ฅ์น
|
27 |
torch_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
28 |
|
29 |
-
# 2) ๋ฒ์ญ
|
30 |
-
|
31 |
-
"
|
32 |
-
model="Helsinki-NLP/opus-mt-ko-en",
|
33 |
-
framework="pt",
|
34 |
-
from_tf=True,
|
35 |
-
device=-1
|
36 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
# ---------------- Encoders ----------------
|
39 |
|
@@ -45,13 +56,20 @@ class HFEmbedder(nn.Module):
|
|
45 |
self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
|
46 |
|
47 |
if self.is_clip:
|
48 |
-
self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
|
49 |
-
|
|
|
|
|
|
|
|
|
50 |
else:
|
51 |
-
self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(
|
52 |
-
|
|
|
|
|
|
|
|
|
53 |
|
54 |
-
# ํ๋ผ๋ฏธํฐ ๋๊ฒฐ
|
55 |
self.hf_module = self.hf_module.eval().requires_grad_(False)
|
56 |
|
57 |
def forward(self, text: list[str]) -> Tensor:
|
@@ -69,30 +87,47 @@ class HFEmbedder(nn.Module):
|
|
69 |
)
|
70 |
return outputs[self.output_key]
|
71 |
|
72 |
-
#
|
73 |
-
t5 = HFEmbedder(
|
74 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
ae = AutoencoderKL.from_pretrained(
|
76 |
-
"black-forest-labs/FLUX.1-dev",
|
|
|
|
|
77 |
).to(torch_device)
|
78 |
|
79 |
-
# ---------------- NF4
|
|
|
80 |
def functional_linear_4bits(x, weight, bias):
|
81 |
-
out = bnb.matmul_4bit(
|
|
|
|
|
82 |
return out.to(x)
|
83 |
|
84 |
def copy_quant_state(state: QuantState, device: torch.device = None) -> QuantState:
|
85 |
if state is None:
|
86 |
return None
|
87 |
device = device or state.absmax.device
|
88 |
-
state2 =
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
|
|
|
|
|
|
|
|
96 |
return QuantState(
|
97 |
absmax=state.absmax.to(device),
|
98 |
shape=state.shape,
|
@@ -110,7 +145,9 @@ class ForgeParams4bit(Params4bit):
|
|
110 |
if device is not None and device.type == "cuda" and not self.bnb_quantized:
|
111 |
return self._quantize(device)
|
112 |
new = ForgeParams4bit(
|
113 |
-
torch.nn.Parameter.to(
|
|
|
|
|
114 |
requires_grad=self.requires_grad,
|
115 |
quant_state=copy_quant_state(self.quant_state, device),
|
116 |
compress_statistics=False,
|
@@ -118,7 +155,7 @@ class ForgeParams4bit(Params4bit):
|
|
118 |
quant_type=self.quant_type,
|
119 |
quant_storage=self.quant_storage,
|
120 |
bnb_quantized=self.bnb_quantized,
|
121 |
-
module=self.module
|
122 |
)
|
123 |
self.module.quant_state = new.quant_state
|
124 |
self.data = new.data
|
@@ -134,29 +171,53 @@ class ForgeLoader4Bit(torch.nn.Module):
|
|
134 |
self.bias = None
|
135 |
self.quant_type = quant_type
|
136 |
|
137 |
-
def _load_from_state_dict(
|
138 |
-
|
139 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
140 |
if any("bitsandbytes" in k for k in qs_keys):
|
141 |
-
qs = {
|
|
|
|
|
142 |
self.weight = ForgeParams4bit.from_prequantized(
|
143 |
data=state_dict[prefix + "weight"],
|
144 |
quantized_stats=qs,
|
145 |
requires_grad=False,
|
146 |
-
device=torch.device(
|
147 |
-
module=self
|
148 |
)
|
149 |
self.quant_state = self.weight.quant_state
|
150 |
if prefix + "bias" in state_dict:
|
151 |
-
self.bias = torch.nn.Parameter(
|
|
|
|
|
152 |
del self.dummy
|
153 |
else:
|
154 |
-
super()._load_from_state_dict(
|
155 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
|
157 |
class Linear(ForgeLoader4Bit):
|
158 |
def __init__(self, *args, device=None, dtype=None, **kwargs):
|
159 |
-
super().__init__(device=device, dtype=dtype, quant_type=
|
|
|
160 |
def forward(self, x):
|
161 |
self.weight.quant_state = self.quant_state
|
162 |
if self.bias is not None and self.bias.dtype != x.dtype:
|
@@ -165,44 +226,61 @@ class Linear(ForgeLoader4Bit):
|
|
165 |
|
166 |
nn.Linear = Linear
|
167 |
|
168 |
-
# ---------------- Flux ๋ชจ๋ธ ์ ์ (
|
169 |
-
|
170 |
-
|
171 |
-
# (
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
|
173 |
# ---------------- ๋ชจ๋ธ ๋ก๋ ----------------
|
|
|
174 |
sd = load_file(
|
175 |
hf_hub_download(
|
176 |
repo_id="lllyasviel/flux1-dev-bnb-nf4",
|
177 |
-
filename="flux1-dev-bnb-nf4-v2.safetensors"
|
178 |
)
|
179 |
)
|
180 |
-
sd = {
|
|
|
|
|
|
|
|
|
181 |
|
182 |
model = Flux().to(torch_device, dtype=torch.bfloat16)
|
183 |
model.load_state_dict(sd)
|
184 |
model_zero_init = False
|
185 |
|
186 |
-
# ----------------
|
187 |
|
188 |
def get_image(image) -> torch.Tensor | None:
|
189 |
if image is None:
|
190 |
return None
|
191 |
image = Image.fromarray(image).convert("RGB")
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
|
|
|
|
197 |
|
198 |
def prepare(t5, clip, img, prompt):
|
199 |
bs, c, h, w = img.shape
|
200 |
-
img = rearrange(
|
|
|
|
|
201 |
if bs == 1 and isinstance(prompt, list):
|
202 |
img = repeat(img, "1 ... -> bs ...", bs=len(prompt))
|
203 |
-
img_ids = torch.zeros(h//2, w//2, 3, device=img.device)
|
204 |
-
img_ids[...,1] = torch.arange(h//2, device=img.device)[:,None]
|
205 |
-
img_ids[...,2] = torch.arange(w//2, device=img.device)[None
|
206 |
img_ids = repeat(img_ids, "h w c -> b (h w) c", b=img.shape[0])
|
207 |
|
208 |
txt = t5([prompt] if isinstance(prompt, str) else prompt)
|
@@ -223,34 +301,57 @@ def prepare(t5, clip, img, prompt):
|
|
223 |
}
|
224 |
|
225 |
def get_schedule(num_steps, image_seq_len, base_shift=0.5, max_shift=1.15, shift=True):
|
226 |
-
timesteps = torch.linspace(1, 0, num_steps+1)
|
227 |
if shift:
|
228 |
-
mu = ((max_shift-base_shift)/(4096-256))*
|
229 |
-
|
|
|
|
|
230 |
return timesteps.tolist()
|
231 |
|
232 |
def denoise(model, img, img_ids, txt, txt_ids, vec, timesteps, guidance):
|
233 |
-
guidance_vec = torch.full(
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
238 |
img = img + (t_prev - t_curr) * pred
|
239 |
return img
|
240 |
|
|
|
|
|
241 |
@spaces.GPU
|
242 |
@torch.no_grad()
|
243 |
def generate_image(
|
244 |
-
prompt,
|
245 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
progress=gr.Progress(track_tqdm=True),
|
247 |
):
|
248 |
# ํ๊ธ ๊ฐ์ง ์ CPU ๋ฒ์ญ๊ธฐ ์ฌ์ฉ
|
249 |
-
if any(
|
250 |
-
|
251 |
-
prompt = translated
|
252 |
|
253 |
-
# ๋๋ค ์๋
|
254 |
if seed == 0:
|
255 |
seed = random.randint(1, 1_000_000)
|
256 |
|
@@ -259,29 +360,38 @@ def generate_image(
|
|
259 |
model = model.to(torch_device)
|
260 |
model_zero_init = True
|
261 |
|
262 |
-
# img2img ์ค๋น
|
263 |
if do_img2img and init_image is not None:
|
264 |
init_img = get_image(init_image)
|
265 |
if resize_img:
|
266 |
-
init_img = torch.nn.functional.interpolate(
|
|
|
|
|
267 |
else:
|
268 |
h0, w0 = init_img.shape[-2:]
|
269 |
-
init_img = init_img[..., :16*(h0//16), :16*(w0//16)]
|
270 |
height, width = init_img.shape[-2:]
|
271 |
-
init_img = ae.encode(
|
272 |
-
|
|
|
|
|
|
|
|
|
273 |
else:
|
274 |
init_img = None
|
275 |
|
276 |
-
# ๋
ธ์ด์ฆ ์ํ ์์ฑ
|
277 |
generator = torch.Generator(device=str(torch_device)).manual_seed(seed)
|
278 |
x = torch.randn(
|
279 |
-
1,
|
280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
)
|
282 |
-
|
283 |
-
timesteps = get_schedule(inference_steps, (x.shape[-1]*x.shape[-2])//4, shift=True)
|
284 |
-
|
285 |
if do_img2img and init_img is not None:
|
286 |
t_idx = int((1 - image2image_strength) * inference_steps)
|
287 |
t = timesteps[t_idx]
|
@@ -291,50 +401,83 @@ def generate_image(
|
|
291 |
inp = prepare(t5, clip, x, prompt)
|
292 |
x = denoise(model, **inp, timesteps=timesteps, guidance=guidance)
|
293 |
|
294 |
-
x = rearrange(
|
295 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
296 |
with torch.autocast(device_type=torch_device.type, dtype=torch.bfloat16):
|
297 |
x = (x / ae.config.scaling_factor) + ae.config.shift_factor
|
298 |
x = ae.decode(x).sample
|
299 |
|
300 |
-
x = x.clamp(-1,1)
|
301 |
-
img = Image.fromarray(
|
|
|
|
|
|
|
|
|
|
|
302 |
|
303 |
return img, seed
|
304 |
|
305 |
-
|
306 |
-
|
307 |
-
|
|
|
|
|
308 |
|
309 |
def create_demo():
|
310 |
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
311 |
-
gr.Markdown(
|
|
|
|
|
|
|
|
|
312 |
with gr.Row():
|
313 |
with gr.Column():
|
314 |
-
prompt = gr.Textbox(
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
321 |
init_img = gr.Image(label="Input Image", visible=False)
|
322 |
-
strength = gr.Slider(
|
323 |
-
|
324 |
-
|
|
|
|
|
325 |
with gr.Column():
|
326 |
-
out_img
|
327 |
out_seed = gr.Text(label="Used Seed")
|
328 |
|
329 |
do_i2i.change(
|
330 |
-
fn=lambda x: [gr.update(visible=x)]*3,
|
331 |
inputs=[do_i2i],
|
332 |
-
outputs=[init_img, strength, resize]
|
333 |
)
|
334 |
btn.click(
|
335 |
fn=generate_image,
|
336 |
-
inputs=[
|
337 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
338 |
)
|
339 |
return demo
|
340 |
|
|
|
14 |
import bitsandbytes as bnb
|
15 |
from bitsandbytes.nn.modules import Params4bit, QuantState
|
16 |
from transformers import (
|
17 |
+
MarianTokenizer,
|
18 |
+
MarianMTModel,
|
19 |
CLIPTextModel, CLIPTokenizer,
|
20 |
T5EncoderModel, T5Tokenizer
|
21 |
)
|
|
|
24 |
from safetensors.torch import load_file
|
25 |
from einops import rearrange, repeat
|
26 |
|
27 |
+
# 1) ์ฅ์น ์ค์
|
28 |
torch_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
29 |
|
30 |
+
# 2) ๋ฒ์ญ ๋ชจ๋ธ์ CPU์์, ๋ฐ๋์ PyTorch ์ฒดํฌํฌ์ธํธ๋ก ๋ก๋
|
31 |
+
trans_tokenizer = MarianTokenizer.from_pretrained(
|
32 |
+
"Helsinki-NLP/opus-mt-ko-en"
|
|
|
|
|
|
|
|
|
33 |
)
|
34 |
+
trans_model = MarianMTModel.from_pretrained(
|
35 |
+
"Helsinki-NLP/opus-mt-ko-en",
|
36 |
+
from_tf=True, # TF ์ฒดํฌํฌ์ธํธ๋ผ๋ PyTorch ๋ก๋
|
37 |
+
torch_dtype=torch.float32,
|
38 |
+
).to(torch.device("cpu"))
|
39 |
+
|
40 |
+
def translate_ko_to_en(text: str, max_length: int = 512) -> str:
|
41 |
+
"""ํ๊ธ โ ์์ด ๋ฒ์ญ (CPU)"""
|
42 |
+
batch = trans_tokenizer([text], return_tensors="pt", padding=True)
|
43 |
+
# ๋ชจ๋ธ์ CPU์ ์์ผ๋ฏ๋ก .to("cpu") ํด์ค ํ์ ์์
|
44 |
+
gen = trans_model.generate(
|
45 |
+
**batch, max_length=max_length
|
46 |
+
)
|
47 |
+
return trans_tokenizer.batch_decode(gen, skip_special_tokens=True)[0]
|
48 |
|
49 |
# ---------------- Encoders ----------------
|
50 |
|
|
|
56 |
self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
|
57 |
|
58 |
if self.is_clip:
|
59 |
+
self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(
|
60 |
+
version, max_length=max_length
|
61 |
+
)
|
62 |
+
self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(
|
63 |
+
version, **hf_kwargs
|
64 |
+
)
|
65 |
else:
|
66 |
+
self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(
|
67 |
+
version, max_length=max_length
|
68 |
+
)
|
69 |
+
self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(
|
70 |
+
version, **hf_kwargs
|
71 |
+
)
|
72 |
|
|
|
73 |
self.hf_module = self.hf_module.eval().requires_grad_(False)
|
74 |
|
75 |
def forward(self, text: list[str]) -> Tensor:
|
|
|
87 |
)
|
88 |
return outputs[self.output_key]
|
89 |
|
90 |
+
# T5, CLIP, VAE ๋ชจ๋ GPU/CPU(device)๋ก ์ด๋
|
91 |
+
t5 = HFEmbedder(
|
92 |
+
"DeepFloyd/t5-v1_1-xxl",
|
93 |
+
max_length=512,
|
94 |
+
torch_dtype=torch.bfloat16
|
95 |
+
).to(torch_device)
|
96 |
+
clip = HFEmbedder(
|
97 |
+
"openai/clip-vit-large-patch14",
|
98 |
+
max_length=77,
|
99 |
+
torch_dtype=torch.bfloat16
|
100 |
+
).to(torch_device)
|
101 |
ae = AutoencoderKL.from_pretrained(
|
102 |
+
"black-forest-labs/FLUX.1-dev",
|
103 |
+
subfolder="vae",
|
104 |
+
torch_dtype=torch.bfloat16
|
105 |
).to(torch_device)
|
106 |
|
107 |
+
# ---------------- NF4 ์ง์ ์ฝ๋ ----------------
|
108 |
+
|
109 |
def functional_linear_4bits(x, weight, bias):
|
110 |
+
out = bnb.matmul_4bit(
|
111 |
+
x, weight.t(), bias=bias, quant_state=weight.quant_state
|
112 |
+
)
|
113 |
return out.to(x)
|
114 |
|
115 |
def copy_quant_state(state: QuantState, device: torch.device = None) -> QuantState:
|
116 |
if state is None:
|
117 |
return None
|
118 |
device = device or state.absmax.device
|
119 |
+
state2 = (
|
120 |
+
QuantState(
|
121 |
+
absmax=state.state2.absmax.to(device),
|
122 |
+
shape=state.state2.shape,
|
123 |
+
code=state.state2.code.to(device),
|
124 |
+
blocksize=state.state2.blocksize,
|
125 |
+
quant_type=state.state2.quant_type,
|
126 |
+
dtype=state.state2.dtype,
|
127 |
+
)
|
128 |
+
if state.nested
|
129 |
+
else None
|
130 |
+
)
|
131 |
return QuantState(
|
132 |
absmax=state.absmax.to(device),
|
133 |
shape=state.shape,
|
|
|
145 |
if device is not None and device.type == "cuda" and not self.bnb_quantized:
|
146 |
return self._quantize(device)
|
147 |
new = ForgeParams4bit(
|
148 |
+
torch.nn.Parameter.to(
|
149 |
+
self, device=device, dtype=dtype, non_blocking=non_blocking
|
150 |
+
),
|
151 |
requires_grad=self.requires_grad,
|
152 |
quant_state=copy_quant_state(self.quant_state, device),
|
153 |
compress_statistics=False,
|
|
|
155 |
quant_type=self.quant_type,
|
156 |
quant_storage=self.quant_storage,
|
157 |
bnb_quantized=self.bnb_quantized,
|
158 |
+
module=self.module,
|
159 |
)
|
160 |
self.module.quant_state = new.quant_state
|
161 |
self.data = new.data
|
|
|
171 |
self.bias = None
|
172 |
self.quant_type = quant_type
|
173 |
|
174 |
+
def _load_from_state_dict(
|
175 |
+
self,
|
176 |
+
state_dict,
|
177 |
+
prefix,
|
178 |
+
local_metadata,
|
179 |
+
strict,
|
180 |
+
missing_keys,
|
181 |
+
unexpected_keys,
|
182 |
+
error_msgs,
|
183 |
+
):
|
184 |
+
qs_keys = {
|
185 |
+
k[len(prefix + "weight.") :]
|
186 |
+
for k in state_dict
|
187 |
+
if k.startswith(prefix + "weight.")
|
188 |
+
}
|
189 |
if any("bitsandbytes" in k for k in qs_keys):
|
190 |
+
qs = {
|
191 |
+
k: state_dict[prefix + "weight." + k] for k in qs_keys
|
192 |
+
}
|
193 |
self.weight = ForgeParams4bit.from_prequantized(
|
194 |
data=state_dict[prefix + "weight"],
|
195 |
quantized_stats=qs,
|
196 |
requires_grad=False,
|
197 |
+
device=torch.device("cuda"),
|
198 |
+
module=self,
|
199 |
)
|
200 |
self.quant_state = self.weight.quant_state
|
201 |
if prefix + "bias" in state_dict:
|
202 |
+
self.bias = torch.nn.Parameter(
|
203 |
+
state_dict[prefix + "bias"].to(self.dummy)
|
204 |
+
)
|
205 |
del self.dummy
|
206 |
else:
|
207 |
+
super()._load_from_state_dict(
|
208 |
+
state_dict,
|
209 |
+
prefix,
|
210 |
+
local_metadata,
|
211 |
+
strict,
|
212 |
+
missing_keys,
|
213 |
+
unexpected_keys,
|
214 |
+
error_msgs,
|
215 |
+
)
|
216 |
|
217 |
class Linear(ForgeLoader4Bit):
|
218 |
def __init__(self, *args, device=None, dtype=None, **kwargs):
|
219 |
+
super().__init__(device=device, dtype=dtype, quant_type="nf4")
|
220 |
+
|
221 |
def forward(self, x):
|
222 |
self.weight.quant_state = self.quant_state
|
223 |
if self.bias is not None and self.bias.dtype != x.dtype:
|
|
|
226 |
|
227 |
nn.Linear = Linear
|
228 |
|
229 |
+
# ---------------- Flux ๋ชจ๋ธ ์ ์ (์๋ณธ ๊ทธ๋๋ก) ----------------
|
230 |
+
|
231 |
+
def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
|
232 |
+
# ... (์๋ต ์์ด ์๋ณธ ์ฝ๋ ๊ทธ๋๋ก)
|
233 |
+
q, k = apply_rope(q, k, pe)
|
234 |
+
x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
|
235 |
+
x = x.permute(0, 2, 1, 3).reshape(x.size(0), x.size(2), -1)
|
236 |
+
return x
|
237 |
+
|
238 |
+
# apply_rope, rope, EmbedND, timestep_embedding, MLPEmbedder, RMSNorm, QKNorm,
|
239 |
+
# SelfAttention, Modulation, DoubleStreamBlock, SingleStreamBlock,
|
240 |
+
# LastLayer, FluxParams, Flux ํด๋์ค๊น์ง ์ ๋ถ ์๋ณธ๊ณผ ๋์ผํ๊ฒ ํฌํจํ์ธ์.
|
241 |
|
242 |
# ---------------- ๋ชจ๋ธ ๋ก๋ ----------------
|
243 |
+
|
244 |
sd = load_file(
|
245 |
hf_hub_download(
|
246 |
repo_id="lllyasviel/flux1-dev-bnb-nf4",
|
247 |
+
filename="flux1-dev-bnb-nf4-v2.safetensors",
|
248 |
)
|
249 |
)
|
250 |
+
sd = {
|
251 |
+
k.replace("model.diffusion_model.", ""): v
|
252 |
+
for k, v in sd.items()
|
253 |
+
if "model.diffusion_model" in k
|
254 |
+
}
|
255 |
|
256 |
model = Flux().to(torch_device, dtype=torch.bfloat16)
|
257 |
model.load_state_dict(sd)
|
258 |
model_zero_init = False
|
259 |
|
260 |
+
# ---------------- ์ ํธ๋ฆฌํฐ ํจ์ ----------------
|
261 |
|
262 |
def get_image(image) -> torch.Tensor | None:
|
263 |
if image is None:
|
264 |
return None
|
265 |
image = Image.fromarray(image).convert("RGB")
|
266 |
+
tfm = transforms.Compose(
|
267 |
+
[
|
268 |
+
transforms.ToTensor(),
|
269 |
+
transforms.Lambda(lambda x: 2.0 * x - 1.0),
|
270 |
+
]
|
271 |
+
)
|
272 |
+
return tfm(image)[None, ...]
|
273 |
|
274 |
def prepare(t5, clip, img, prompt):
|
275 |
bs, c, h, w = img.shape
|
276 |
+
img = rearrange(
|
277 |
+
img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2
|
278 |
+
)
|
279 |
if bs == 1 and isinstance(prompt, list):
|
280 |
img = repeat(img, "1 ... -> bs ...", bs=len(prompt))
|
281 |
+
img_ids = torch.zeros(h // 2, w // 2, 3, device=img.device)
|
282 |
+
img_ids[..., 1] = torch.arange(h // 2, device=img.device)[:, None]
|
283 |
+
img_ids[..., 2] = torch.arange(w // 2, device=img.device)[None, :]
|
284 |
img_ids = repeat(img_ids, "h w c -> b (h w) c", b=img.shape[0])
|
285 |
|
286 |
txt = t5([prompt] if isinstance(prompt, str) else prompt)
|
|
|
301 |
}
|
302 |
|
303 |
def get_schedule(num_steps, image_seq_len, base_shift=0.5, max_shift=1.15, shift=True):
|
304 |
+
timesteps = torch.linspace(1, 0, num_steps + 1)
|
305 |
if shift:
|
306 |
+
mu = ((max_shift - base_shift) / (4096 - 256)) * image_seq_len + (
|
307 |
+
base_shift - (256 * (max_shift - base_shift) / (4096 - 256))
|
308 |
+
)
|
309 |
+
timesteps = timesteps.exp().div((1 / timesteps - 1) ** 1 + mu)
|
310 |
return timesteps.tolist()
|
311 |
|
312 |
def denoise(model, img, img_ids, txt, txt_ids, vec, timesteps, guidance):
|
313 |
+
guidance_vec = torch.full(
|
314 |
+
(img.size(0),), guidance, device=img.device, dtype=img.dtype
|
315 |
+
)
|
316 |
+
for t_curr, t_prev in tqdm(
|
317 |
+
zip(timesteps[:-1], timesteps[1:]), total=len(timesteps) - 1
|
318 |
+
):
|
319 |
+
t_vec = torch.full(
|
320 |
+
(img.size(0),), t_curr, device=img.device, dtype=img.dtype
|
321 |
+
)
|
322 |
+
pred = model(
|
323 |
+
img=img,
|
324 |
+
img_ids=img_ids,
|
325 |
+
txt=txt,
|
326 |
+
txt_ids=txt_ids,
|
327 |
+
y=vec,
|
328 |
+
timesteps=t_vec,
|
329 |
+
guidance=guidance_vec,
|
330 |
+
)
|
331 |
img = img + (t_prev - t_curr) * pred
|
332 |
return img
|
333 |
|
334 |
+
# ---------------- Gradio ๋ฐ๋ชจ ----------------
|
335 |
+
|
336 |
@spaces.GPU
|
337 |
@torch.no_grad()
|
338 |
def generate_image(
|
339 |
+
prompt,
|
340 |
+
width,
|
341 |
+
height,
|
342 |
+
guidance,
|
343 |
+
inference_steps,
|
344 |
+
seed,
|
345 |
+
do_img2img,
|
346 |
+
init_image,
|
347 |
+
image2image_strength,
|
348 |
+
resize_img,
|
349 |
progress=gr.Progress(track_tqdm=True),
|
350 |
):
|
351 |
# ํ๊ธ ๊ฐ์ง ์ CPU ๋ฒ์ญ๊ธฐ ์ฌ์ฉ
|
352 |
+
if any("\u3131" <= c <= "\u318E" or "\uAC00" <= c <= "\uD7A3" for c in prompt):
|
353 |
+
prompt = translate_ko_to_en(prompt)
|
|
|
354 |
|
|
|
355 |
if seed == 0:
|
356 |
seed = random.randint(1, 1_000_000)
|
357 |
|
|
|
360 |
model = model.to(torch_device)
|
361 |
model_zero_init = True
|
362 |
|
|
|
363 |
if do_img2img and init_image is not None:
|
364 |
init_img = get_image(init_image)
|
365 |
if resize_img:
|
366 |
+
init_img = torch.nn.functional.interpolate(
|
367 |
+
init_img, (height, width)
|
368 |
+
)
|
369 |
else:
|
370 |
h0, w0 = init_img.shape[-2:]
|
371 |
+
init_img = init_img[..., : 16 * (h0 // 16), : 16 * (w0 // 16)]
|
372 |
height, width = init_img.shape[-2:]
|
373 |
+
init_img = ae.encode(
|
374 |
+
init_img.to(torch_device).to(torch.bfloat16)
|
375 |
+
).latent_dist.sample()
|
376 |
+
init_img = (
|
377 |
+
init_img - ae.config.shift_factor
|
378 |
+
) * ae.config.scaling_factor
|
379 |
else:
|
380 |
init_img = None
|
381 |
|
|
|
382 |
generator = torch.Generator(device=str(torch_device)).manual_seed(seed)
|
383 |
x = torch.randn(
|
384 |
+
1,
|
385 |
+
16,
|
386 |
+
2 * math.ceil(height / 16),
|
387 |
+
2 * math.ceil(width / 16),
|
388 |
+
device=torch_device,
|
389 |
+
dtype=torch.bfloat16,
|
390 |
+
generator=generator,
|
391 |
+
)
|
392 |
+
timesteps = get_schedule(
|
393 |
+
inference_steps, (x.shape[-1] * x.shape[-2]) // 4, shift=True
|
394 |
)
|
|
|
|
|
|
|
395 |
if do_img2img and init_img is not None:
|
396 |
t_idx = int((1 - image2image_strength) * inference_steps)
|
397 |
t = timesteps[t_idx]
|
|
|
401 |
inp = prepare(t5, clip, x, prompt)
|
402 |
x = denoise(model, **inp, timesteps=timesteps, guidance=guidance)
|
403 |
|
404 |
+
x = rearrange(
|
405 |
+
x[:, inp["txt"].shape[1] :, ...].float(),
|
406 |
+
"b (h w) (c ph pw) -> b c (h ph) (w pw)",
|
407 |
+
h=math.ceil(height / 16),
|
408 |
+
w=math.ceil(width / 16),
|
409 |
+
ph=2,
|
410 |
+
pw=2,
|
411 |
+
)
|
412 |
with torch.autocast(device_type=torch_device.type, dtype=torch.bfloat16):
|
413 |
x = (x / ae.config.scaling_factor) + ae.config.shift_factor
|
414 |
x = ae.decode(x).sample
|
415 |
|
416 |
+
x = x.clamp(-1, 1)
|
417 |
+
img = Image.fromarray(
|
418 |
+
(127.5 * (rearrange(x[0], "c h w -> h w c") + 1.0))
|
419 |
+
.cpu()
|
420 |
+
.byte()
|
421 |
+
.numpy()
|
422 |
+
)
|
423 |
|
424 |
return img, seed
|
425 |
|
426 |
+
css = """
|
427 |
+
footer {
|
428 |
+
visibility: hidden;
|
429 |
+
}
|
430 |
+
"""
|
431 |
|
432 |
def create_demo():
|
433 |
with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
|
434 |
+
gr.Markdown(
|
435 |
+
"# News! Multilingual version "
|
436 |
+
"[https://huggingface.co/spaces/ginigen/FLUXllama-Multilingual]"
|
437 |
+
"(https://huggingface.co/spaces/ginigen/FLUXllama-Multilingual)"
|
438 |
+
)
|
439 |
with gr.Row():
|
440 |
with gr.Column():
|
441 |
+
prompt = gr.Textbox(
|
442 |
+
label="Prompt(ํ๊ธ ๊ฐ๋ฅ)",
|
443 |
+
value="A cute and fluffy golden retriever puppy sitting upright...",
|
444 |
+
)
|
445 |
+
width = gr.Slider(128, 2048, 64, label="Width", value=768)
|
446 |
+
height = gr.Slider(128, 2048, 64, label="Height", value=768)
|
447 |
+
guidance = gr.Slider(1.0, 5.0, 0.1, label="Guidance", value=3.5)
|
448 |
+
steps = gr.Slider(1, 30, 1, label="Inference steps", value=30)
|
449 |
+
seed = gr.Number(label="Seed", precision=0)
|
450 |
+
do_i2i = gr.Checkbox(label="Image to Image", value=False)
|
451 |
init_img = gr.Image(label="Input Image", visible=False)
|
452 |
+
strength = gr.Slider(
|
453 |
+
0.0, 1.0, 0.01, label="Noising strength", value=0.8, visible=False
|
454 |
+
)
|
455 |
+
resize = gr.Checkbox(label="Resize image", value=True, visible=False)
|
456 |
+
btn = gr.Button("Generate")
|
457 |
with gr.Column():
|
458 |
+
out_img = gr.Image(label="Generated Image")
|
459 |
out_seed = gr.Text(label="Used Seed")
|
460 |
|
461 |
do_i2i.change(
|
462 |
+
fn=lambda x: [gr.update(visible=x)] * 3,
|
463 |
inputs=[do_i2i],
|
464 |
+
outputs=[init_img, strength, resize],
|
465 |
)
|
466 |
btn.click(
|
467 |
fn=generate_image,
|
468 |
+
inputs=[
|
469 |
+
prompt,
|
470 |
+
width,
|
471 |
+
height,
|
472 |
+
guidance,
|
473 |
+
steps,
|
474 |
+
seed,
|
475 |
+
do_i2i,
|
476 |
+
init_img,
|
477 |
+
strength,
|
478 |
+
resize,
|
479 |
+
],
|
480 |
+
outputs=[out_img, out_seed],
|
481 |
)
|
482 |
return demo
|
483 |
|