File size: 14,414 Bytes
040ea64 |
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 |
from typing import Any, Dict, List, Optional, Union
import numpy as np
import torch
import torch.nn as nn
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import PeftAdapterMixin, FromOriginalModelMixin
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormContinuous
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.models.embeddings import TimestepEmbedding, get_timestep_embedding
from diffusers.models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
# Support different diffusers versions
try:
from diffusers.models.embeddings import FluxPosEmbed as EmbedND
except:
from diffusers.models.transformers.transformer_flux import rope
class EmbedND(nn.Module):
def __init__(self, theta: int, axes_dim: List[int]):
super().__init__()
self.theta = theta
self.axes_dim = axes_dim
def forward(self, ids: torch.Tensor) -> torch.Tensor:
n_axes = ids.shape[-1]
emb = torch.cat(
[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
dim=-3,
)
return emb.unsqueeze(1)
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class Timesteps(nn.Module):
def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float, scale: int = 1,time_theta=10000):
super().__init__()
self.num_channels = num_channels
self.flip_sin_to_cos = flip_sin_to_cos
self.downscale_freq_shift = downscale_freq_shift
self.scale = scale
self.time_theta=time_theta
def forward(self, timesteps):
t_emb = get_timestep_embedding(
timesteps,
self.num_channels,
flip_sin_to_cos=self.flip_sin_to_cos,
downscale_freq_shift=self.downscale_freq_shift,
scale=self.scale,
max_period=self.time_theta
)
return t_emb
class TimestepProjEmbeddings(nn.Module):
def __init__(self, embedding_dim, time_theta):
super().__init__()
self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0,time_theta=time_theta)
self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
def forward(self, timestep, dtype):
timesteps_proj = self.time_proj(timestep)
timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=dtype)) # (N, D)
return timesteps_emb
"""
Based on FluxPipeline with several changes:
- no pooled embeddings
- We use zero padding for prompts
- No guidance embedding since this is not a distilled version
"""
class BriaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
"""
The Transformer model introduced in Flux.
Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
Parameters:
patch_size (`int`): Patch size to turn the input data into small patches.
in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
"""
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
patch_size: int = 1,
in_channels: int = 64,
num_layers: int = 19,
num_single_layers: int = 38,
attention_head_dim: int = 128,
num_attention_heads: int = 24,
joint_attention_dim: int = 4096,
pooled_projection_dim: int = None,
guidance_embeds: bool = False,
axes_dims_rope: List[int] = [16, 56, 56],
rope_theta = 10000,
time_theta = 10000
):
super().__init__()
self.out_channels = in_channels
self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
self.pos_embed = EmbedND(theta=rope_theta, axes_dim=axes_dims_rope)
self.time_embed = TimestepProjEmbeddings(
embedding_dim=self.inner_dim,time_theta=time_theta
)
# if pooled_projection_dim:
# self.pooled_text_embed = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim=self.inner_dim, act_fn="silu")
if guidance_embeds:
self.guidance_embed = TimestepProjEmbeddings(embedding_dim=self.inner_dim)
self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)
self.transformer_blocks = nn.ModuleList(
[
FluxTransformerBlock(
dim=self.inner_dim,
num_attention_heads=self.config.num_attention_heads,
attention_head_dim=self.config.attention_head_dim,
)
for i in range(self.config.num_layers)
]
)
self.single_transformer_blocks = nn.ModuleList(
[
FluxSingleTransformerBlock(
dim=self.inner_dim,
num_attention_heads=self.config.num_attention_heads,
attention_head_dim=self.config.attention_head_dim,
)
for i in range(self.config.num_single_layers)
]
)
self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
self.gradient_checkpointing = False
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
def forward(
self,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor = None,
pooled_projections: torch.Tensor = None,
timestep: torch.LongTensor = None,
img_ids: torch.Tensor = None,
txt_ids: torch.Tensor = None,
guidance: torch.Tensor = None,
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
return_dict: bool = True,
controlnet_block_samples = None,
controlnet_single_block_samples=None,
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
"""
The [`FluxTransformer2DModel`] forward method.
Args:
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
Input `hidden_states`.
encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
from the embeddings of input conditions.
timestep ( `torch.LongTensor`):
Used to indicate denoising step.
block_controlnet_hidden_states: (`list` of `torch.Tensor`):
A list of tensors that if specified are added to the residuals of transformer blocks.
joint_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
tuple.
Returns:
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
`tuple` where the first element is the sample tensor.
"""
if joint_attention_kwargs is not None:
joint_attention_kwargs = joint_attention_kwargs.copy()
lora_scale = joint_attention_kwargs.pop("scale", 1.0)
else:
lora_scale = 1.0
if USE_PEFT_BACKEND:
# weight the lora layers by setting `lora_scale` for each PEFT layer
scale_lora_layers(self, lora_scale)
else:
if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
logger.warning(
"Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
)
hidden_states = self.x_embedder(hidden_states)
timestep = timestep.to(hidden_states.dtype)
if guidance is not None:
guidance = guidance.to(hidden_states.dtype)
else:
guidance = None
# temb = (
# self.time_text_embed(timestep, pooled_projections)
# if guidance is None
# else self.time_text_embed(timestep, guidance, pooled_projections)
# )
temb = self.time_embed(timestep,dtype=hidden_states.dtype)
# if pooled_projections:
# temb+=self.pooled_text_embed(pooled_projections)
if guidance:
temb+=self.guidance_embed(guidance,dtype=hidden_states.dtype)
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
if len(txt_ids.shape)==2:
ids = torch.cat((txt_ids, img_ids), dim=0)
else:
ids = torch.cat((txt_ids, img_ids), dim=1)
image_rotary_emb = self.pos_embed(ids)
for index_block, block in enumerate(self.transformer_blocks):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(block),
hidden_states,
encoder_hidden_states,
temb,
image_rotary_emb,
**ckpt_kwargs,
)
else:
encoder_hidden_states, hidden_states = block(
hidden_states=hidden_states,
encoder_hidden_states=encoder_hidden_states,
temb=temb,
image_rotary_emb=image_rotary_emb,
)
# controlnet residual
if controlnet_block_samples is not None:
interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
interval_control = int(np.ceil(interval_control))
hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
for index_block, block in enumerate(self.single_transformer_blocks):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(block),
hidden_states,
temb,
image_rotary_emb,
**ckpt_kwargs,
)
else:
hidden_states = block(
hidden_states=hidden_states,
temb=temb,
image_rotary_emb=image_rotary_emb,
)
# controlnet residual
if controlnet_single_block_samples is not None:
interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
interval_control = int(np.ceil(interval_control))
hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+ controlnet_single_block_samples[index_block // interval_control]
)
hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
hidden_states = self.norm_out(hidden_states, temb)
output = self.proj_out(hidden_states)
if USE_PEFT_BACKEND:
# remove `lora_scale` from each PEFT layer
unscale_lora_layers(self, lora_scale)
if not return_dict:
return (output,)
return Transformer2DModelOutput(sample=output)
|