File size: 17,538 Bytes
e2aa741 |
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 |
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
import einops
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from .layers import LLamaFeedForward, RMSNorm
def modulate(x, scale):
return x * (1 + scale)
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size, frequency_embedding_size=256):
super().__init__()
self.hidden_size = hidden_size
self.frequency_embedding_size = frequency_embedding_size
self.mlp = nn.Sequential(
nn.Linear(self.frequency_embedding_size, self.hidden_size),
nn.SiLU(),
nn.Linear(self.hidden_size, self.hidden_size),
)
@staticmethod
def timestep_embedding(t, dim, max_period=10000):
"""
Create sinusoidal timestep embeddings.
:param t: a 1-D Tensor of N indices, one per batch element.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an (N, D) Tensor of positional embeddings.
"""
half = dim // 2
freqs = torch.exp(
-np.log(max_period) * torch.arange(0, half, dtype=t.dtype) / half
).to(t.device)
args = t[:, :, None] * freqs[None, :]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :, :1])], dim=-1)
return embedding
def forward(self, t):
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
t_freq = t_freq.to(self.mlp[0].weight.dtype)
return self.mlp(t_freq)
class FinalLayer(nn.Module):
def __init__(self, hidden_size, num_patches, out_channels):
super().__init__()
self.norm_final = nn.LayerNorm(hidden_size, eps=1e-6, elementwise_affine=False)
self.linear = nn.Linear(hidden_size, num_patches * out_channels)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(),
nn.Linear(min(hidden_size, 1024), hidden_size),
)
def forward(self, x, c):
scale = self.adaLN_modulation(c)
x = modulate(self.norm_final(x), scale)
x = self.linear(x)
return x
class Attention(nn.Module):
def __init__(
self,
dim,
n_heads,
n_kv_heads=None,
qk_norm=False,
y_dim=0,
base_seqlen=None,
proportional_attn=False,
attention_dropout=0.0,
max_position_embeddings=384,
):
super().__init__()
self.dim = dim
self.n_heads = n_heads
self.n_kv_heads = n_kv_heads or n_heads
self.qk_norm = qk_norm
self.y_dim = y_dim
self.base_seqlen = base_seqlen
self.proportional_attn = proportional_attn
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.head_dim = dim // n_heads
self.wq = nn.Linear(dim, n_heads * self.head_dim, bias=False)
self.wk = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias=False)
self.wv = nn.Linear(dim, self.n_kv_heads * self.head_dim, bias=False)
if y_dim > 0:
self.wk_y = nn.Linear(y_dim, self.n_kv_heads * self.head_dim, bias=False)
self.wv_y = nn.Linear(y_dim, self.n_kv_heads * self.head_dim, bias=False)
self.gate = nn.Parameter(torch.zeros(n_heads))
self.wo = nn.Linear(n_heads * self.head_dim, dim, bias=False)
if qk_norm:
self.q_norm = nn.LayerNorm(self.n_heads * self.head_dim)
self.k_norm = nn.LayerNorm(self.n_kv_heads * self.head_dim)
if y_dim > 0:
self.ky_norm = nn.LayerNorm(self.n_kv_heads * self.head_dim, eps=1e-6)
else:
self.ky_norm = nn.Identity()
else:
self.q_norm = nn.Identity()
self.k_norm = nn.Identity()
self.ky_norm = nn.Identity()
@staticmethod
def apply_rotary_emb(xq, xk, freqs_cis):
# xq, xk: [batch_size, seq_len, n_heads, head_dim]
# freqs_cis: [1, seq_len, 1, head_dim]
xq_ = xq.float().reshape(*xq.shape[:-1], -1, 2)
xk_ = xk.float().reshape(*xk.shape[:-1], -1, 2)
xq_complex = torch.view_as_complex(xq_)
xk_complex = torch.view_as_complex(xk_)
freqs_cis = freqs_cis.unsqueeze(2)
# Apply freqs_cis
xq_out = xq_complex * freqs_cis
xk_out = xk_complex * freqs_cis
# Convert back to real numbers
xq_out = torch.view_as_real(xq_out).flatten(-2)
xk_out = torch.view_as_real(xk_out).flatten(-2)
return xq_out.type_as(xq), xk_out.type_as(xk)
def forward(
self,
x,
x_mask,
freqs_cis,
y=None,
y_mask=None,
init_cache=False,
):
bsz, seqlen, _ = x.size()
xq = self.wq(x)
xk = self.wk(x)
xv = self.wv(x)
if x_mask is None:
x_mask = torch.ones(bsz, seqlen, dtype=torch.bool, device=x.device)
inp_dtype = xq.dtype
xq = self.q_norm(xq)
xk = self.k_norm(xk)
xq = xq.view(bsz, seqlen, self.n_heads, self.head_dim)
xk = xk.view(bsz, seqlen, self.n_kv_heads, self.head_dim)
xv = xv.view(bsz, seqlen, self.n_kv_heads, self.head_dim)
if self.n_kv_heads != self.n_heads:
n_rep = self.n_heads // self.n_kv_heads
xk = xk.repeat_interleave(n_rep, dim=2)
xv = xv.repeat_interleave(n_rep, dim=2)
freqs_cis = freqs_cis.to(xq.device)
xq, xk = self.apply_rotary_emb(xq, xk, freqs_cis)
output = (
F.scaled_dot_product_attention(
xq.permute(0, 2, 1, 3),
xk.permute(0, 2, 1, 3),
xv.permute(0, 2, 1, 3),
attn_mask=x_mask.bool().view(bsz, 1, 1, seqlen).expand(-1, self.n_heads, seqlen, -1),
scale=None,
)
.permute(0, 2, 1, 3)
.to(inp_dtype)
)
if hasattr(self, "wk_y"):
yk = self.ky_norm(self.wk_y(y)).view(bsz, -1, self.n_kv_heads, self.head_dim)
yv = self.wv_y(y).view(bsz, -1, self.n_kv_heads, self.head_dim)
n_rep = self.n_heads // self.n_kv_heads
# if n_rep >= 1:
# yk = yk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
# yv = yv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
if n_rep >= 1:
yk = einops.repeat(yk, "b l h d -> b l (repeat h) d", repeat=n_rep)
yv = einops.repeat(yv, "b l h d -> b l (repeat h) d", repeat=n_rep)
output_y = F.scaled_dot_product_attention(
xq.permute(0, 2, 1, 3),
yk.permute(0, 2, 1, 3),
yv.permute(0, 2, 1, 3),
y_mask.view(bsz, 1, 1, -1).expand(bsz, self.n_heads, seqlen, -1).to(torch.bool),
).permute(0, 2, 1, 3)
output_y = output_y * self.gate.tanh().view(1, 1, -1, 1)
output = output + output_y
output = output.flatten(-2)
output = self.wo(output)
return output.to(inp_dtype)
class TransformerBlock(nn.Module):
"""
Corresponds to the Transformer block in the JAX code.
"""
def __init__(
self,
dim,
n_heads,
n_kv_heads,
multiple_of,
ffn_dim_multiplier,
norm_eps,
qk_norm,
y_dim,
max_position_embeddings,
):
super().__init__()
self.attention = Attention(dim, n_heads, n_kv_heads, qk_norm, y_dim=y_dim, max_position_embeddings=max_position_embeddings)
self.feed_forward = LLamaFeedForward(
dim=dim,
hidden_dim=4 * dim,
multiple_of=multiple_of,
ffn_dim_multiplier=ffn_dim_multiplier,
)
self.attention_norm1 = RMSNorm(dim, eps=norm_eps)
self.attention_norm2 = RMSNorm(dim, eps=norm_eps)
self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(),
nn.Linear(min(dim, 1024), 4 * dim),
)
self.attention_y_norm = RMSNorm(y_dim, eps=norm_eps)
def forward(
self,
x,
x_mask,
freqs_cis,
y,
y_mask,
adaln_input=None,
):
if adaln_input is not None:
scales_gates = self.adaLN_modulation(adaln_input)
# TODO: Duong - check the dimension of chunking
# scale_msa, gate_msa, scale_mlp, gate_mlp = scales_gates.chunk(4, dim=-1)
scale_msa, gate_msa, scale_mlp, gate_mlp = scales_gates.chunk(4, dim=-1)
x = x + torch.tanh(gate_msa) * self.attention_norm2(
self.attention(
modulate(self.attention_norm1(x), scale_msa), # ok
x_mask,
freqs_cis,
self.attention_y_norm(y), # ok
y_mask,
)
)
x = x + torch.tanh(gate_mlp) * self.ffn_norm2(
self.feed_forward(
modulate(self.ffn_norm1(x), scale_mlp),
)
)
else:
x = x + self.attention_norm2(
self.attention(
self.attention_norm1(x),
x_mask,
freqs_cis,
self.attention_y_norm(y),
y_mask,
)
)
x = x + self.ffn_norm2(self.feed_forward(self.ffn_norm1(x)))
return x
class NextDiT(ModelMixin, ConfigMixin):
"""
Diffusion model with a Transformer backbone for joint image-video training.
"""
@register_to_config
def __init__(
self,
input_size=(1, 32, 32),
patch_size=(1, 2, 2),
in_channels=16,
hidden_size=4096,
depth=32,
num_heads=32,
num_kv_heads=None,
multiple_of=256,
ffn_dim_multiplier=None,
norm_eps=1e-5,
pred_sigma=False,
caption_channels=4096,
qk_norm=False,
norm_type="rms",
model_max_length=120,
rotary_max_length=384,
rotary_max_length_t=None
):
super().__init__()
self.input_size = input_size
self.patch_size = patch_size
self.in_channels = in_channels
self.hidden_size = hidden_size
self.depth = depth
self.num_heads = num_heads
self.num_kv_heads = num_kv_heads or num_heads
self.multiple_of = multiple_of
self.ffn_dim_multiplier = ffn_dim_multiplier
self.norm_eps = norm_eps
self.pred_sigma = pred_sigma
self.caption_channels = caption_channels
self.qk_norm = qk_norm
self.norm_type = norm_type
self.model_max_length = model_max_length
self.rotary_max_length = rotary_max_length
self.rotary_max_length_t = rotary_max_length_t
self.out_channels = in_channels * 2 if pred_sigma else in_channels
self.x_embedder = nn.Linear(np.prod(self.patch_size) * in_channels, hidden_size)
self.t_embedder = TimestepEmbedder(min(hidden_size, 1024))
self.y_embedder = nn.Sequential(
nn.LayerNorm(caption_channels, eps=1e-6),
nn.Linear(caption_channels, min(hidden_size, 1024)),
)
self.layers = nn.ModuleList([
TransformerBlock(
dim=hidden_size,
n_heads=num_heads,
n_kv_heads=self.num_kv_heads,
multiple_of=multiple_of,
ffn_dim_multiplier=ffn_dim_multiplier,
norm_eps=norm_eps,
qk_norm=qk_norm,
y_dim=caption_channels,
max_position_embeddings=rotary_max_length,
)
for _ in range(depth)
])
self.final_layer = FinalLayer(
hidden_size=hidden_size,
num_patches=np.prod(patch_size),
out_channels=self.out_channels,
)
assert (hidden_size // num_heads) % 6 == 0, "3d rope needs head dim to be divisible by 6"
self.freqs_cis = self.precompute_freqs_cis(
hidden_size // num_heads,
self.rotary_max_length,
end_t=self.rotary_max_length_t
)
def to(self, *args, **kwargs):
self = super().to(*args, **kwargs)
# self.freqs_cis = self.freqs_cis.to(*args, **kwargs)
return self
@staticmethod
def precompute_freqs_cis(
dim: int,
end: int,
end_t: int = None,
theta: float = 10000.0,
scale_factor: float = 1.0,
scale_watershed: float = 1.0,
timestep: float = 1.0,
):
if timestep < scale_watershed:
linear_factor = scale_factor
ntk_factor = 1.0
else:
linear_factor = 1.0
ntk_factor = scale_factor
theta = theta * ntk_factor
freqs = 1.0 / (theta ** (torch.arange(0, dim, 6)[: (dim // 6)] / dim)) / linear_factor
timestep = torch.arange(end, dtype=torch.float32)
freqs = torch.outer(timestep, freqs).float()
freqs_cis = torch.exp(1j * freqs)
if end_t is not None:
freqs_t = 1.0 / (theta ** (torch.arange(0, dim, 6)[: (dim // 6)] / dim)) / linear_factor
timestep_t = torch.arange(end_t, dtype=torch.float32)
freqs_t = torch.outer(timestep_t, freqs_t).float()
freqs_cis_t = torch.exp(1j * freqs_t)
freqs_cis_t = freqs_cis_t.view(end_t, 1, 1, dim // 6).repeat(1, end, end, 1)
else:
end_t = end
freqs_cis_t = freqs_cis.view(end_t, 1, 1, dim // 6).repeat(1, end, end, 1)
freqs_cis_h = freqs_cis.view(1, end, 1, dim // 6).repeat(end_t, 1, end, 1)
freqs_cis_w = freqs_cis.view(1, 1, end, dim // 6).repeat(end_t, end, 1, 1)
freqs_cis = torch.cat([freqs_cis_t, freqs_cis_h, freqs_cis_w], dim=-1).view(end_t, end, end, -1)
return freqs_cis
def forward(
self,
samples,
timesteps,
encoder_hidden_states,
encoder_attention_mask,
scale_factor: float = 1.0, # scale_factor for rotary embedding
scale_watershed: float = 1.0, # scale_watershed for rotary embedding
):
if samples.ndim == 4: # B C H W
samples = samples[:, None, ...] # B F C H W
precomputed_freqs_cis = None
if scale_factor != 1 or scale_watershed != 1:
precomputed_freqs_cis = self.precompute_freqs_cis(
self.hidden_size // self.num_heads,
self.rotary_max_length,
end_t=self.rotary_max_length_t,
scale_factor=scale_factor,
scale_watershed=scale_watershed,
timestep=torch.max(timesteps.cpu()).item()
)
if len(timesteps.shape) == 5:
t, *_ = self.patchify(timesteps, precomputed_freqs_cis)
timesteps = t.mean(dim=-1)
elif len(timesteps.shape) == 1:
timesteps = timesteps[:, None, None, None, None].expand_as(samples)
t, *_ = self.patchify(timesteps, precomputed_freqs_cis)
timesteps = t.mean(dim=-1)
samples, T, H, W, freqs_cis = self.patchify(samples, precomputed_freqs_cis)
samples = self.x_embedder(samples)
t = self.t_embedder(timesteps)
encoder_attention_mask_float = encoder_attention_mask[..., None].float()
encoder_hidden_states_pool = (encoder_hidden_states * encoder_attention_mask_float).sum(dim=1) / (encoder_attention_mask_float.sum(dim=1) + 1e-8)
encoder_hidden_states_pool = encoder_hidden_states_pool.to(samples.dtype)
y = self.y_embedder(encoder_hidden_states_pool)
y = y.unsqueeze(1).expand(-1, samples.size(1), -1)
adaln_input = t + y
for block in self.layers:
samples = block(samples, None, freqs_cis, encoder_hidden_states, encoder_attention_mask, adaln_input)
samples = self.final_layer(samples, adaln_input)
samples = self.unpatchify(samples, T, H, W)
return samples
def patchify(self, x, precompute_freqs_cis=None):
# pytorch is C, H, W
B, T, C, H, W = x.size()
pT, pH, pW = self.patch_size
x = x.view(B, T // pT, pT, C, H // pH, pH, W // pW, pW)
x = x.permute(0, 1, 4, 6, 2, 5, 7, 3)
x = x.reshape(B, -1, pT * pH * pW * C)
if precompute_freqs_cis is None:
freqs_cis = self.freqs_cis[: T // pT, :H // pH, :W // pW].reshape(-1, * self.freqs_cis.shape[3:])[None].to(x.device)
else:
freqs_cis = precompute_freqs_cis[: T // pT, :H // pH, :W // pW].reshape(-1, * precompute_freqs_cis.shape[3:])[None].to(x.device)
return x, T // pT, H // pH, W // pW, freqs_cis
def unpatchify(self, x, T, H, W):
B = x.size(0)
C = self.out_channels
pT, pH, pW = self.patch_size
x = x.view(B, T, H, W, pT, pH, pW, C)
x = x.permute(0, 1, 4, 7, 2, 5, 3, 6)
x = x.reshape(B, T * pT, C, H * pH, W * pW)
return x
|