Spaces:
Running
on
L40S
Running
on
L40S
File size: 24,169 Bytes
5aa9bba |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 |
# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Any, Dict, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
from diffusers.models.attention import FeedForward
from diffusers.models.attention_processor import (
Attention,
AttentionProcessor,
FluxAttnProcessor2_0,
FusedFluxAttnProcessor2_0,
)
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
from diffusers.models.modeling_outputs import Transformer2DModelOutput
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@maybe_allow_in_graph
class FluxSingleTransformerBlock(nn.Module):
r"""
A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
Reference: https://arxiv.org/abs/2403.03206
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
processing of `context` conditions.
"""
def __init__(self, dim, num_attention_heads, attention_head_dim, mlp_ratio=4.0):
super().__init__()
self.mlp_hidden_dim = int(dim * mlp_ratio)
self.norm = AdaLayerNormZeroSingle(dim)
self.proj_mlp = nn.Linear(dim, self.mlp_hidden_dim)
self.act_mlp = nn.GELU(approximate="tanh")
self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
processor = FluxAttnProcessor2_0()
self.attn = Attention(
query_dim=dim,
cross_attention_dim=None,
dim_head=attention_head_dim,
heads=num_attention_heads,
out_dim=dim,
bias=True,
processor=processor,
qk_norm="rms_norm",
eps=1e-6,
pre_only=True,
)
def forward(
self,
hidden_states: torch.FloatTensor,
temb: torch.FloatTensor,
image_emb=None,
image_rotary_emb=None,
):
residual = hidden_states
norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
attn_output = self.attn(
hidden_states=norm_hidden_states,
image_rotary_emb=image_rotary_emb,
image_emb=image_emb,
)
hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
gate = gate.unsqueeze(1) # torch.Size([1, 1, 3072])
hidden_states = gate * self.proj_out(hidden_states) # torch.Size([1, 4352, 3072])
hidden_states = residual + hidden_states
if hidden_states.dtype == torch.float16:
hidden_states = hidden_states.clip(-65504, 65504)
return hidden_states
@maybe_allow_in_graph
class FluxTransformerBlock(nn.Module):
r"""
A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
Reference: https://arxiv.org/abs/2403.03206
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
processing of `context` conditions.
"""
def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm="rms_norm", eps=1e-6):
super().__init__()
self.norm1 = AdaLayerNormZero(dim)
self.norm1_context = AdaLayerNormZero(dim)
if hasattr(F, "scaled_dot_product_attention"):
processor = FluxAttnProcessor2_0()
else:
raise ValueError(
"The current PyTorch version does not support the `scaled_dot_product_attention` function."
)
self.attn = Attention(
query_dim=dim,
cross_attention_dim=None,
added_kv_proj_dim=dim,
dim_head=attention_head_dim,
heads=num_attention_heads,
out_dim=dim,
context_pre_only=False,
bias=True,
processor=processor,
qk_norm=qk_norm,
eps=eps,
)
self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
# let chunk size default to None
self._chunk_size = None
self._chunk_dim = 0
def forward(
self,
hidden_states: torch.FloatTensor,
encoder_hidden_states: torch.FloatTensor,
temb: torch.FloatTensor,
image_emb=None,
image_rotary_emb=None,
):
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
encoder_hidden_states, emb=temb
)
# Attention.
attn_output, context_attn_output = self.attn(
hidden_states=norm_hidden_states,
encoder_hidden_states=norm_encoder_hidden_states,
image_rotary_emb=image_rotary_emb,
image_emb=image_emb,
)
# Process attention outputs for the `hidden_states`.
attn_output = gate_msa.unsqueeze(1) * attn_output
hidden_states = hidden_states + attn_output
norm_hidden_states = self.norm2(hidden_states)
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
ff_output = self.ff(norm_hidden_states)
ff_output = gate_mlp.unsqueeze(1) * ff_output
hidden_states = hidden_states + ff_output
# Process attention outputs for the `encoder_hidden_states`.
context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
encoder_hidden_states = encoder_hidden_states + context_attn_output
norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
context_ff_output = self.ff_context(norm_encoder_hidden_states)
encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
if encoder_hidden_states.dtype == torch.float16:
encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
return encoder_hidden_states, hidden_states
class FluxTransformer2DModel(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
_no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
@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 = 768,
guidance_embeds: bool = False,
axes_dims_rope: Tuple[int] = (16, 56, 56),
):
super().__init__()
self.out_channels = in_channels
self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
text_time_guidance_cls = (
CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
)
self.time_text_embed = text_time_guidance_cls(
embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_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
@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()
for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
return processors
for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)
return processors
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.
Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.
"""
count = len(self.attn_processors.keys())
if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))
for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
def fuse_qkv_projections(self):
"""
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
are fused. For cross-attention modules, key and value projection matrices are fused.
<Tip warning={true}>
This API is 🧪 experimental.
</Tip>
"""
self.original_attn_processors = None
for _, attn_processor in self.attn_processors.items():
if "Added" in str(attn_processor.__class__.__name__):
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
self.original_attn_processors = self.attn_processors
for module in self.modules():
if isinstance(module, Attention):
module.fuse_projections(fuse=True)
self.set_attn_processor(FusedFluxAttnProcessor2_0())
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
def unfuse_qkv_projections(self):
"""Disables the fused QKV projection if enabled.
<Tip warning={true}>
This API is 🧪 experimental.
</Tip>
"""
if self.original_attn_processors is not None:
self.set_attn_processor(self.original_attn_processors)
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,
image_emb: torch.FloatTensor = 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,
controlnet_block_samples=None,
controlnet_single_block_samples=None,
return_dict: bool = True,
) -> 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) * 1000
if guidance is not None:
guidance = guidance.to(hidden_states.dtype) * 1000
else:
guidance = None
temb = (
self.time_text_embed(timestep, pooled_projections)
if guidance is None
else self.time_text_embed(timestep, guidance, pooled_projections)
)
# torch.Size([1, 512*num_prompt, 4096]) -> torch.Size([1, 512*num_prompt, 3072])
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
if txt_ids.ndim == 3:
logger.warning(
"Passing `txt_ids` 3d torch.Tensor is deprecated."
"Please remove the batch dimension and pass it as a 2d torch Tensor"
)
txt_ids = txt_ids[0]
if img_ids.ndim == 3:
logger.warning(
"Passing `img_ids` 3d torch.Tensor is deprecated."
"Please remove the batch dimension and pass it as a 2d torch Tensor"
)
img_ids = img_ids[0]
ids = torch.cat((txt_ids, img_ids), dim=0)
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_emb,
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_emb=image_emb,
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_emb,
image_rotary_emb,
**ckpt_kwargs,
)
else:
hidden_states = block(
hidden_states=hidden_states,
temb=temb,
image_emb=image_emb,
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) |