Spaces:
Running on Zero
Running on Zero
File size: 9,606 Bytes
3bd5145 | 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 | import torch, warnings, glob, os
import numpy as np
from PIL import Image
from einops import repeat, reduce
from typing import Optional, Union
from dataclasses import dataclass
import numpy as np
from PIL import Image
from typing import Optional
class BasePipeline(torch.nn.Module):
def __init__(
self,
device="cuda", torch_dtype=torch.float16,
height_division_factor=64, width_division_factor=64,
time_division_factor=None, time_division_remainder=None,
):
super().__init__()
# The device and torch_dtype is used for the storage of intermediate variables, not models.
self.device = device
self.torch_dtype = torch_dtype
# The following parameters are used for shape check.
self.height_division_factor = height_division_factor
self.width_division_factor = width_division_factor
self.time_division_factor = time_division_factor
self.time_division_remainder = time_division_remainder
self.vram_management_enabled = False
def to(self, *args, **kwargs):
device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
if device is not None:
self.device = device
if dtype is not None:
self.torch_dtype = dtype
super().to(*args, **kwargs)
return self
def check_resize_height_width(self, height, width, num_frames=None):
# Shape check
if height % self.height_division_factor != 0:
height = (height + self.height_division_factor - 1) // self.height_division_factor * self.height_division_factor
print(f"height % {self.height_division_factor} != 0. We round it up to {height}.")
if width % self.width_division_factor != 0:
width = (width + self.width_division_factor - 1) // self.width_division_factor * self.width_division_factor
print(f"width % {self.width_division_factor} != 0. We round it up to {width}.")
if num_frames is None:
return height, width
else:
if num_frames % self.time_division_factor != self.time_division_remainder:
num_frames = (num_frames + self.time_division_factor - 1) // self.time_division_factor * self.time_division_factor + self.time_division_remainder
print(f"num_frames % {self.time_division_factor} != {self.time_division_remainder}. We round it up to {num_frames}.")
return height, width, num_frames
def preprocess_image(self, image, torch_dtype=None, device=None, pattern="B C H W", min_value=-1, max_value=1):
# Transform a PIL.Image to torch.Tensor
image = torch.Tensor(np.array(image, dtype=np.float32))
image = image.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
image = image * ((max_value - min_value) / 255) + min_value
image = repeat(image, f"H W C -> {pattern}", **({"B": 1} if "B" in pattern else {}))
return image
def preprocess_video(self, video, torch_dtype=None, device=None, pattern="B C T H W", min_value=-1, max_value=1):
# Transform a list of PIL.Image to torch.Tensor
video = [self.preprocess_image(image, torch_dtype=torch_dtype, device=device, min_value=min_value, max_value=max_value) for image in video]
video = torch.stack(video, dim=pattern.index("T") // 2)
return video
def vae_output_to_image(self, vae_output, pattern="B C H W", min_value=-1, max_value=1):
# Transform a torch.Tensor to PIL.Image
if pattern != "H W C":
vae_output = reduce(vae_output, f"{pattern} -> H W C", reduction="mean")
image = ((vae_output - min_value) * (255 / (max_value - min_value))).clip(0, 255)
image = image.to(device="cpu", dtype=torch.uint8)
image = Image.fromarray(image.numpy())
return image
def vae_output_to_video(self, vae_output, pattern="B C T H W", min_value=-1, max_value=1):
# Transform a torch.Tensor to list of PIL.Image
if pattern != "T H W C":
vae_output = reduce(vae_output, f"{pattern} -> T H W C", reduction="mean")
video = [self.vae_output_to_image(image, pattern="H W C", min_value=min_value, max_value=max_value) for image in vae_output]
return video
def load_models_to_device(self, model_names=[]):
if self.vram_management_enabled:
# offload models
for name, model in self.named_children():
if name not in model_names:
if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
for module in model.modules():
if hasattr(module, "offload"):
module.offload()
else:
model.cpu()
torch.cuda.empty_cache()
# onload models
for name, model in self.named_children():
if name in model_names:
if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
for module in model.modules():
if hasattr(module, "onload"):
module.onload()
else:
model.to(self.device)
def generate_noise(self, shape, seed=None, rand_device="cpu", rand_torch_dtype=torch.float32, device=None, torch_dtype=None):
# Initialize Gaussian noise
generator = None if seed is None else torch.Generator(rand_device).manual_seed(seed)
noise = torch.randn(shape, generator=generator, device=rand_device, dtype=rand_torch_dtype)
noise = noise.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
return noise
def enable_cpu_offload(self):
warnings.warn("`enable_cpu_offload` will be deprecated. Please use `enable_vram_management`.")
self.vram_management_enabled = True
def get_vram(self):
return torch.cuda.mem_get_info(self.device)[1] / (1024 ** 3)
def freeze_except(self, model_names):
for name, model in self.named_children():
if name in model_names:
model.train()
model.requires_grad_(True)
else:
model.eval()
model.requires_grad_(False)
def blend_with_mask(self, base, addition, mask):
return base * (1 - mask) + addition * mask
def step(self, scheduler, latents, progress_id, noise_pred, input_latents=None, inpaint_mask=None, **kwargs):
timestep = scheduler.timesteps[progress_id]
if inpaint_mask is not None:
noise_pred_expected = scheduler.return_to_timestep(scheduler.timesteps[progress_id], latents, input_latents)
noise_pred = self.blend_with_mask(noise_pred_expected, noise_pred, inpaint_mask)
latents_next = scheduler.step(noise_pred, timestep, latents)
return latents_next
class PipelineUnit:
def __init__(
self,
seperate_cfg: bool = False,
take_over: bool = False,
input_params: tuple[str] = None,
input_params_posi: dict[str, str] = None,
input_params_nega: dict[str, str] = None,
onload_model_names: tuple[str] = None
):
self.seperate_cfg = seperate_cfg
self.take_over = take_over
self.input_params = input_params
self.input_params_posi = input_params_posi
self.input_params_nega = input_params_nega
self.onload_model_names = onload_model_names
def process(self, pipe: BasePipeline, inputs: dict, positive=True, **kwargs) -> dict:
raise NotImplementedError("`process` is not implemented.")
class PipelineUnitRunner:
def __init__(self):
pass
def __call__(self, unit: PipelineUnit, pipe: BasePipeline, inputs_shared: dict, inputs_posi: dict, inputs_nega: dict) -> tuple[dict, dict]:
if unit.take_over:
# Let the pipeline unit take over this function.
inputs_shared, inputs_posi, inputs_nega = unit.process(pipe, inputs_shared=inputs_shared, inputs_posi=inputs_posi, inputs_nega=inputs_nega)
elif unit.seperate_cfg:
# Positive side
processor_inputs = {name: inputs_posi.get(name_) for name, name_ in unit.input_params_posi.items()}
if unit.input_params is not None:
for name in unit.input_params:
processor_inputs[name] = inputs_shared.get(name)
processor_outputs = unit.process(pipe, **processor_inputs)
inputs_posi.update(processor_outputs)
# Negative side
if inputs_shared["cfg_scale"] != 1:
processor_inputs = {name: inputs_nega.get(name_) for name, name_ in unit.input_params_nega.items()}
if unit.input_params is not None:
for name in unit.input_params:
processor_inputs[name] = inputs_shared.get(name)
processor_outputs = unit.process(pipe, **processor_inputs)
inputs_nega.update(processor_outputs)
else:
inputs_nega.update(processor_outputs)
else:
processor_inputs = {name: inputs_shared.get(name) for name in unit.input_params}
processor_outputs = unit.process(pipe, **processor_inputs)
inputs_shared.update(processor_outputs)
return inputs_shared, inputs_posi, inputs_nega
|