Spaces:
Runtime error
Runtime error
File size: 19,174 Bytes
a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 7dc6a72 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 8eae2b8 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 a0a2ed9 7dc6a72 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 |
import gc
import os
import traceback
from typing import List, Literal, Optional, Union
import numpy as np
import torch
from diffusers import AutoencoderTiny, StableDiffusionPipeline
from PIL import Image
from polygraphy import cuda
from streamdiffusion import StreamDiffusion
from streamdiffusion.image_utils import postprocess_image
torch.set_grad_enabled(False)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
class StreamDiffusionWrapper:
def __init__(
self,
model_id: str,
t_index_list: List[int],
mode: Literal["img2img", "txt2img"] = "img2img",
output_type: Literal["pil", "pt", "np", "latent"] = "pil",
lcm_lora_id: Optional[str] = None,
vae_id: Optional[str] = None,
device: Literal["cpu", "cuda"] = "cuda",
dtype: torch.dtype = torch.float16,
frame_buffer_size: int = 1,
width: int = 512,
height: int = 512,
warmup: int = 10,
acceleration: Literal["none", "xformers", "sfast", "tensorrt"] = "xformers",
is_drawing: bool = True,
device_ids: Optional[List[int]] = None,
use_lcm_lora: bool = True,
use_tiny_vae: bool = True,
enable_similar_image_filter: bool = False,
similar_image_filter_threshold: float = 0.98,
use_denoising_batch: bool = True,
cfg_type: Literal["none", "full", "self", "initialize"] = "none",
use_safety_checker: bool = False,
):
if mode == "txt2img":
if cfg_type != "none":
raise ValueError(
f"txt2img mode accepts only cfg_type = 'none', but got {cfg_type}"
)
if use_denoising_batch and frame_buffer_size > 1:
raise ValueError(
"txt2img mode cannot use denoising batch with frame_buffer_size > 1."
)
if mode == "img2img":
if not use_denoising_batch:
raise NotImplementedError(
"img2img mode must use denoising batch for now."
)
self.sd_turbo = "turbo" in model_id
self.device = device
self.dtype = dtype
self.width = width
self.height = height
self.mode = mode
self.output_type = output_type
self.frame_buffer_size = frame_buffer_size
self.batch_size = (
len(t_index_list) * frame_buffer_size
if use_denoising_batch
else frame_buffer_size
)
self.use_denoising_batch = use_denoising_batch
self.use_safety_checker = use_safety_checker
self.stream = self._load_model(
model_id=model_id,
lcm_lora_id=lcm_lora_id,
vae_id=vae_id,
t_index_list=t_index_list,
acceleration=acceleration,
warmup=warmup,
is_drawing=is_drawing,
use_lcm_lora=use_lcm_lora,
use_tiny_vae=use_tiny_vae,
cfg_type=cfg_type,
)
if device_ids is not None:
self.stream.unet = torch.nn.DataParallel(
self.stream.unet, device_ids=device_ids
)
if enable_similar_image_filter:
self.stream.enable_similar_image_filter(similar_image_filter_threshold)
def prepare(
self,
prompt: str,
negative_prompt: str = "",
num_inference_steps: int = 50,
guidance_scale: float = 1.2,
delta: float = 1.0,
) -> None:
"""
Prepares the model for inference.
Parameters
----------
prompt : str
The prompt to generate images from.
num_inference_steps : int, optional
The number of inference steps to perform, by default 50.
"""
self.stream.prepare(
prompt,
negative_prompt,
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
delta=delta,
)
def __call__(
self,
image: Optional[Union[str, Image.Image, torch.Tensor]] = None,
prompt: Optional[str] = None,
) -> Union[Image.Image, List[Image.Image]]:
"""
Performs img2img or txt2img based on the mode.
Parameters
----------
image : Optional[Union[str, Image.Image, torch.Tensor]]
The image to generate from.
prompt : Optional[str]
The prompt to generate images from.
Returns
-------
Union[Image.Image, List[Image.Image]]
The generated image.
"""
if self.mode == "img2img":
return self.img2img(image)
else:
return self.txt2img(prompt)
def txt2img(
self, prompt: str
) -> Union[Image.Image, List[Image.Image], torch.Tensor, np.ndarray]:
"""
Performs txt2img.
Parameters
----------
prompt : str
The prompt to generate images from.
Returns
-------
Union[Image.Image, List[Image.Image]]
The generated image.
"""
self.stream.update_prompt(prompt)
if self.sd_turbo:
image_tensor = self.stream.txt2img_sd_turbo(self.batch_size)
else:
image_tensor = self.stream.txt2img(self.frame_buffer_size)
image = self.postprocess_image(image_tensor, output_type=self.output_type)
if self.use_safety_checker:
safety_checker_input = self.feature_extractor(
image, return_tensors="pt"
).to(self.device)
_, has_nsfw_concept = self.safety_checker(
images=image_tensor.to(self.dtype),
clip_input=safety_checker_input.pixel_values.to(self.dtype),
)
image = self.nsfw_fallback_img if has_nsfw_concept[0] else image
return image
def img2img(
self, image: Union[str, Image.Image, torch.Tensor]
) -> Union[Image.Image, List[Image.Image], torch.Tensor, np.ndarray]:
"""
Performs img2img.
Parameters
----------
image : Union[str, Image.Image, torch.Tensor]
The image to generate from.
Returns
-------
Image.Image
The generated image.
"""
if isinstance(image, str) or isinstance(image, Image.Image):
image = self.preprocess_image(image)
image_tensor = self.stream(image)
return self.postprocess_image(image_tensor, output_type=self.output_type)
def preprocess_image(self, image: Union[str, Image.Image]) -> torch.Tensor:
"""
Preprocesses the image.
Parameters
----------
image : Union[str, Image.Image, torch.Tensor]
The image to preprocess.
Returns
-------
torch.Tensor
The preprocessed image.
"""
if isinstance(image, str):
image = Image.open(image).convert("RGB").resize((self.width, self.height))
if isinstance(image, Image.Image):
image = image.convert("RGB").resize((self.width, self.height))
return self.stream.image_processor.preprocess(
image, self.height, self.width
).to(device=self.device, dtype=self.dtype)
def postprocess_image(
self, image_tensor: torch.Tensor, output_type: str = "pil"
) -> Union[Image.Image, List[Image.Image], torch.Tensor, np.ndarray]:
"""
Postprocesses the image.
Parameters
----------
image_tensor : torch.Tensor
The image tensor to postprocess.
Returns
-------
Union[Image.Image, List[Image.Image]]
The postprocessed image.
"""
if self.frame_buffer_size > 1:
return postprocess_image(image_tensor.cpu(), output_type=output_type)
else:
return postprocess_image(image_tensor.cpu(), output_type=output_type)[0]
def _load_model(
self,
model_id: str,
t_index_list: List[int],
lcm_lora_id: Optional[str] = None,
vae_id: Optional[str] = None,
acceleration: Literal["none", "sfast", "tensorrt"] = "tensorrt",
is_drawing: bool = True,
warmup: int = 10,
use_lcm_lora: bool = True,
use_tiny_vae: bool = True,
cfg_type: Literal["none", "full", "self", "initialize"] = "self",
):
"""
Loads the model.
This method does the following:
1. Loads the model from the model_id.
2. Loads and fuses the LCM-LoRA model from the lcm_lora_id if needed.
3. Loads the VAE model from the vae_id if needed.
4. Enables acceleration if needed.
5. Prepares the model for inference.
6. Warms up the model.
Parameters
----------
model_id : str
The model id to load.
t_index_list : List[int]
The t_index_list to use for inference.
lcm_lora_id : Optional[str], optional
The lcm_lora_id to load, by default None.
vae_id : Optional[str], optional
The vae_id to load, by default None.
acceleration : Literal["none", "xfomers", "sfast", "tensorrt"], optional
The acceleration method to use, by default "tensorrt".
warmup : int, optional
The number of warmup steps to perform, by default 10.
is_drawing : bool, optional
Whether to draw the image or not, by default True.
use_lcm_lora : bool, optional
Whether to use LCM-LoRA or not, by default True.
use_tiny_vae : bool, optional
Whether to use TinyVAE or not, by default True.
cfg_type : Literal["none", "full", "self", "initialize"], optional
The cfg_type to use, by default "self".
"""
try: # Load from local directory
pipe: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
model_id,
).to(device=self.device, dtype=self.dtype)
except ValueError: # Load from huggingface
pipe: StableDiffusionPipeline = StableDiffusionPipeline.from_single_file(
model_id
).to(device=self.device, dtype=self.dtype)
except Exception: # No model found
traceback.print_exc()
print("Model load has failed. Doesn't exist.")
exit()
if self.use_safety_checker:
from transformers import CLIPFeatureExtractor
from diffusers.pipelines.stable_diffusion.safety_checker import (
StableDiffusionSafetyChecker,
)
self.safety_checker = StableDiffusionSafetyChecker.from_pretrained(
"CompVis/stable-diffusion-safety-checker"
).to(pipe.device)
self.feature_extractor = CLIPFeatureExtractor.from_pretrained(
"openai/clip-vit-base-patch32"
)
self.nsfw_fallback_img = Image.new("RGB", (512, 512), (0, 0, 0))
stream = StreamDiffusion(
pipe=pipe,
t_index_list=t_index_list,
torch_dtype=self.dtype,
width=self.width,
height=self.height,
is_drawing=is_drawing,
frame_buffer_size=self.frame_buffer_size,
use_denoising_batch=self.use_denoising_batch,
cfg_type=cfg_type,
)
if not self.sd_turbo:
if use_lcm_lora:
if lcm_lora_id is not None:
stream.load_lcm_lora(
pretrained_model_name_or_path_or_dict=lcm_lora_id
)
else:
stream.load_lcm_lora()
stream.fuse_lora()
if use_tiny_vae:
if vae_id is not None:
stream.vae = AutoencoderTiny.from_pretrained(vae_id).to(
device=pipe.device, dtype=pipe.dtype
)
else:
stream.vae = AutoencoderTiny.from_pretrained("madebyollin/taesd").to(
device=pipe.device, dtype=pipe.dtype
)
try:
if acceleration == "xformers":
stream.pipe.enable_xformers_memory_efficient_attention()
if acceleration == "tensorrt":
from streamdiffusion.acceleration.tensorrt import (
TorchVAEEncoder,
compile_unet,
compile_vae_decoder,
compile_vae_encoder,
)
from streamdiffusion.acceleration.tensorrt.engine import (
AutoencoderKLEngine,
UNet2DConditionModelEngine,
)
from streamdiffusion.acceleration.tensorrt.models import (
VAE,
UNet,
VAEEncoder,
)
def create_prefix(
max_batch_size: int,
min_batch_size: int,
):
return f"{model_id}--lcm_lora-{use_tiny_vae}--tiny_vae-{use_lcm_lora}--max_batch-{max_batch_size}--min_batch-{min_batch_size}--mode-{self.mode}"
engine_dir = os.path.join("engines")
unet_path = os.path.join(
engine_dir,
create_prefix(
stream.trt_unet_batch_size, stream.trt_unet_batch_size
),
"unet.engine",
)
vae_encoder_path = os.path.join(
engine_dir,
create_prefix(
self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
),
"vae_encoder.engine",
)
vae_decoder_path = os.path.join(
engine_dir,
create_prefix(
self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
),
"vae_decoder.engine",
)
if not os.path.exists(unet_path):
os.makedirs(os.path.dirname(unet_path), exist_ok=True)
unet_model = UNet(
fp16=True,
device=stream.device,
max_batch_size=stream.trt_unet_batch_size,
min_batch_size=stream.trt_unet_batch_size,
embedding_dim=stream.text_encoder.config.hidden_size,
unet_dim=stream.unet.config.in_channels,
)
compile_unet(
stream.unet,
unet_model,
unet_path + ".onnx",
unet_path + ".opt.onnx",
unet_path,
opt_batch_size=stream.trt_unet_batch_size,
)
if not os.path.exists(vae_decoder_path):
os.makedirs(os.path.dirname(vae_decoder_path), exist_ok=True)
stream.vae.forward = stream.vae.decode
vae_decoder_model = VAE(
device=stream.device,
max_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
min_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
)
compile_vae_decoder(
stream.vae,
vae_decoder_model,
vae_decoder_path + ".onnx",
vae_decoder_path + ".opt.onnx",
vae_decoder_path,
opt_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
)
delattr(stream.vae, "forward")
if not os.path.exists(vae_encoder_path):
os.makedirs(os.path.dirname(vae_encoder_path), exist_ok=True)
vae_encoder = TorchVAEEncoder(stream.vae).to(torch.device("cuda"))
vae_encoder_model = VAEEncoder(
device=stream.device,
max_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
min_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
)
compile_vae_encoder(
vae_encoder,
vae_encoder_model,
vae_encoder_path + ".onnx",
vae_encoder_path + ".opt.onnx",
vae_encoder_path,
opt_batch_size=self.batch_size
if self.mode == "txt2img"
else stream.frame_bff_size,
)
cuda_steram = cuda.Stream()
vae_config = stream.vae.config
vae_dtype = stream.vae.dtype
stream.unet = UNet2DConditionModelEngine(
unet_path, cuda_steram, use_cuda_graph=False
)
stream.vae = AutoencoderKLEngine(
vae_encoder_path,
vae_decoder_path,
cuda_steram,
stream.pipe.vae_scale_factor,
use_cuda_graph=False,
)
setattr(stream.vae, "config", vae_config)
setattr(stream.vae, "dtype", vae_dtype)
gc.collect()
torch.cuda.empty_cache()
print("TensorRT acceleration enabled.")
if acceleration == "sfast":
from streamdiffusion.acceleration.sfast import (
accelerate_with_stable_fast,
)
stream = accelerate_with_stable_fast(stream)
print("StableFast acceleration enabled.")
except Exception:
traceback.print_exc()
print("Acceleration has failed. Falling back to normal mode.")
stream.prepare(
"",
"",
num_inference_steps=50,
guidance_scale=1.1
if stream.cfg_type in ["full", "self", "initialize"]
else 1.0,
generator=torch.manual_seed(2),
)
return stream
|