Spaces:
Running
on
Zero
Running
on
Zero
Update models/model.py
Browse files- models/model.py +318 -198
models/model.py
CHANGED
|
@@ -1,25 +1,9 @@
|
|
| 1 |
-
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
-
# All rights reserved.
|
| 3 |
-
|
| 4 |
-
# This source code is licensed under the license found in the
|
| 5 |
-
# LICENSE file in the root directory of this source tree.
|
| 6 |
-
# --------------------------------------------------------
|
| 7 |
-
# References:
|
| 8 |
-
# GLIDE: https://github.com/openai/glide-text2im
|
| 9 |
-
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
|
| 10 |
-
# --------------------------------------------------------
|
| 11 |
-
|
| 12 |
import functools
|
| 13 |
import logging
|
| 14 |
import math
|
| 15 |
from typing import Optional, Tuple, List
|
| 16 |
|
| 17 |
-
# from apex.normalization import FusedRMSNorm as RMSNorm
|
| 18 |
from .components import RMSNorm
|
| 19 |
-
import fairscale.nn.model_parallel.initialize as fs_init
|
| 20 |
-
from fairscale.nn.model_parallel.layers import (
|
| 21 |
-
ColumnParallelLinear, RowParallelLinear, ParallelEmbedding,
|
| 22 |
-
)
|
| 23 |
from flash_attn import flash_attn_varlen_func
|
| 24 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
| 25 |
import torch
|
|
@@ -38,22 +22,25 @@ def modulate(x, shift, scale):
|
|
| 38 |
# Embedding Layers for Timesteps and Class Labels #
|
| 39 |
#############################################################################
|
| 40 |
|
|
|
|
| 41 |
class ParallelTimestepEmbedder(nn.Module):
|
| 42 |
"""
|
| 43 |
Embeds scalar timesteps into vector representations.
|
| 44 |
"""
|
|
|
|
| 45 |
def __init__(self, hidden_size, frequency_embedding_size=256):
|
| 46 |
super().__init__()
|
| 47 |
self.mlp = nn.Sequential(
|
| 48 |
-
|
| 49 |
-
frequency_embedding_size,
|
| 50 |
-
|
| 51 |
-
|
| 52 |
),
|
| 53 |
nn.SiLU(),
|
| 54 |
-
|
| 55 |
-
hidden_size,
|
| 56 |
-
|
|
|
|
| 57 |
),
|
| 58 |
)
|
| 59 |
self.frequency_embedding_size = frequency_embedding_size
|
|
@@ -71,16 +58,16 @@ class ParallelTimestepEmbedder(nn.Module):
|
|
| 71 |
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
| 72 |
half = dim // 2
|
| 73 |
freqs = torch.exp(
|
| 74 |
-
-math.log(max_period)
|
| 75 |
-
|
| 76 |
-
|
| 77 |
).to(device=t.device)
|
| 78 |
args = t[:, None].float() * freqs[None]
|
| 79 |
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
| 80 |
if dim % 2:
|
| 81 |
-
embedding = torch.cat(
|
| 82 |
-
embedding, torch.zeros_like(embedding[:, :1])
|
| 83 |
-
|
| 84 |
return embedding
|
| 85 |
|
| 86 |
def forward(self, t):
|
|
@@ -93,12 +80,13 @@ class ParallelLabelEmbedder(nn.Module):
|
|
| 93 |
r"""Embeds class labels into vector representations. Also handles label
|
| 94 |
dropout for classifier-free guidance.
|
| 95 |
"""
|
|
|
|
| 96 |
def __init__(self, num_classes, hidden_size, dropout_prob):
|
| 97 |
super().__init__()
|
| 98 |
use_cfg_embedding = int(dropout_prob > 0)
|
| 99 |
-
self.embedding_table =
|
| 100 |
-
num_classes + use_cfg_embedding,
|
| 101 |
-
|
| 102 |
)
|
| 103 |
self.num_classes = num_classes
|
| 104 |
self.dropout_prob = dropout_prob
|
|
@@ -108,15 +96,10 @@ class ParallelLabelEmbedder(nn.Module):
|
|
| 108 |
Drops labels to enable classifier-free guidance.
|
| 109 |
"""
|
| 110 |
if force_drop_ids is None:
|
| 111 |
-
drop_ids =
|
| 112 |
-
labels.shape[0], device=labels.device
|
| 113 |
-
) < self.dropout_prob
|
| 114 |
-
drop_ids = drop_ids.cuda()
|
| 115 |
-
dist.broadcast(
|
| 116 |
-
drop_ids,
|
| 117 |
-
fs_init.get_model_parallel_src_rank(),
|
| 118 |
-
fs_init.get_model_parallel_group(),
|
| 119 |
)
|
|
|
|
| 120 |
drop_ids = drop_ids.to(labels.device)
|
| 121 |
else:
|
| 122 |
drop_ids = force_drop_ids == 1
|
|
@@ -132,13 +115,21 @@ class ParallelLabelEmbedder(nn.Module):
|
|
| 132 |
|
| 133 |
|
| 134 |
#############################################################################
|
| 135 |
-
# Core
|
| 136 |
#############################################################################
|
| 137 |
|
| 138 |
|
| 139 |
class Attention(nn.Module):
|
| 140 |
"""Multi-head attention module."""
|
| 141 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 142 |
"""
|
| 143 |
Initialize the Attention module.
|
| 144 |
|
|
@@ -150,38 +141,44 @@ class Attention(nn.Module):
|
|
| 150 |
"""
|
| 151 |
super().__init__()
|
| 152 |
self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
|
| 153 |
-
model_parallel_size =
|
| 154 |
self.n_local_heads = n_heads // model_parallel_size
|
| 155 |
self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
|
| 156 |
self.n_rep = self.n_local_heads // self.n_local_kv_heads
|
| 157 |
self.head_dim = dim // n_heads
|
| 158 |
|
| 159 |
-
self.wq =
|
| 160 |
-
dim,
|
| 161 |
-
|
|
|
|
| 162 |
)
|
| 163 |
-
self.wk =
|
| 164 |
-
dim,
|
| 165 |
-
|
|
|
|
| 166 |
)
|
| 167 |
-
self.wv =
|
| 168 |
-
dim,
|
| 169 |
-
|
|
|
|
| 170 |
)
|
| 171 |
if y_dim > 0:
|
| 172 |
-
self.wk_y =
|
| 173 |
-
y_dim,
|
| 174 |
-
|
|
|
|
| 175 |
)
|
| 176 |
-
self.wv_y =
|
| 177 |
-
y_dim,
|
| 178 |
-
|
|
|
|
| 179 |
)
|
| 180 |
self.gate = nn.Parameter(torch.zeros([self.n_local_heads]))
|
| 181 |
|
| 182 |
-
self.wo =
|
| 183 |
-
n_heads * self.head_dim,
|
| 184 |
-
|
|
|
|
| 185 |
)
|
| 186 |
|
| 187 |
if qk_norm:
|
|
@@ -194,7 +191,7 @@ class Attention(nn.Module):
|
|
| 194 |
else:
|
| 195 |
self.q_norm = self.k_norm = nn.Identity()
|
| 196 |
self.ky_norm = nn.Identity()
|
| 197 |
-
|
| 198 |
# for proportional attention computation
|
| 199 |
self.base_seqlen = None
|
| 200 |
self.proportional_attn = False
|
|
@@ -224,8 +221,7 @@ class Attention(nn.Module):
|
|
| 224 |
ndim = x.ndim
|
| 225 |
assert 0 <= 1 < ndim
|
| 226 |
assert freqs_cis.shape == (x.shape[1], x.shape[-1])
|
| 227 |
-
shape = [d if i == 1 or i == ndim - 1 else 1
|
| 228 |
-
for i, d in enumerate(x.shape)]
|
| 229 |
return freqs_cis.view(*shape)
|
| 230 |
|
| 231 |
@staticmethod
|
|
@@ -259,13 +255,17 @@ class Attention(nn.Module):
|
|
| 259 |
return x_out.type_as(x_in)
|
| 260 |
|
| 261 |
# copied from huggingface modeling_llama.py
|
| 262 |
-
def _upad_input(
|
|
|
|
|
|
|
| 263 |
|
| 264 |
def _get_unpad_data(attention_mask):
|
| 265 |
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
| 266 |
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
| 267 |
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
| 268 |
-
cu_seqlens = F.pad(
|
|
|
|
|
|
|
| 269 |
return (
|
| 270 |
indices,
|
| 271 |
cu_seqlens,
|
|
@@ -276,14 +276,19 @@ class Attention(nn.Module):
|
|
| 276 |
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
| 277 |
|
| 278 |
key_layer = index_first_axis(
|
| 279 |
-
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
|
|
|
|
| 280 |
)
|
| 281 |
value_layer = index_first_axis(
|
| 282 |
-
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
|
|
|
|
| 283 |
)
|
| 284 |
if query_length == kv_seq_len:
|
| 285 |
query_layer = index_first_axis(
|
| 286 |
-
query_layer.reshape(
|
|
|
|
|
|
|
|
|
|
| 287 |
)
|
| 288 |
cu_seqlens_q = cu_seqlens_k
|
| 289 |
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
|
@@ -298,7 +303,9 @@ class Attention(nn.Module):
|
|
| 298 |
else:
|
| 299 |
# The -q_len: slice assumes left padding.
|
| 300 |
attention_mask = attention_mask[:, -query_length:]
|
| 301 |
-
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(
|
|
|
|
|
|
|
| 302 |
|
| 303 |
return (
|
| 304 |
query_layer,
|
|
@@ -347,15 +354,22 @@ class Attention(nn.Module):
|
|
| 347 |
|
| 348 |
if dtype in [torch.float16, torch.bfloat16]:
|
| 349 |
# begin var_len flash attn
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 353 |
|
| 354 |
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
| 355 |
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
| 356 |
|
| 357 |
if self.proportional_attn:
|
| 358 |
-
softmax_scale = math.sqrt(
|
|
|
|
|
|
|
| 359 |
else:
|
| 360 |
softmax_scale = math.sqrt(1 / self.head_dim)
|
| 361 |
|
|
@@ -367,24 +381,32 @@ class Attention(nn.Module):
|
|
| 367 |
cu_seqlens_k=cu_seqlens_k,
|
| 368 |
max_seqlen_q=max_seqlen_in_batch_q,
|
| 369 |
max_seqlen_k=max_seqlen_in_batch_k,
|
| 370 |
-
dropout_p=0
|
| 371 |
causal=False,
|
| 372 |
-
softmax_scale=softmax_scale
|
| 373 |
)
|
| 374 |
output = pad_input(attn_output_unpad, indices_q, bsz, seqlen)
|
| 375 |
# end var_len_flash_attn
|
| 376 |
|
| 377 |
else:
|
| 378 |
-
output =
|
| 379 |
-
|
| 380 |
-
|
| 381 |
-
|
| 382 |
-
|
| 383 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 384 |
|
| 385 |
if hasattr(self, "wk_y"):
|
| 386 |
# todo better flash_attn support
|
| 387 |
-
yk = self.ky_norm(self.wk_y(y)).view(
|
|
|
|
|
|
|
| 388 |
yv = self.wv_y(y).view(bsz, -1, self.n_local_kv_heads, self.head_dim)
|
| 389 |
n_rep = self.n_local_heads // self.n_local_kv_heads
|
| 390 |
if n_rep >= 1:
|
|
@@ -394,7 +416,7 @@ class Attention(nn.Module):
|
|
| 394 |
xq.permute(0, 2, 1, 3),
|
| 395 |
yk.permute(0, 2, 1, 3),
|
| 396 |
yv.permute(0, 2, 1, 3),
|
| 397 |
-
y_mask.view(bsz, 1, 1, -1).expand(bsz, self.n_local_heads, seqlen, -1)
|
| 398 |
).permute(0, 2, 1, 3)
|
| 399 |
output_y = output_y * self.gate.tanh().view(1, 1, -1, 1)
|
| 400 |
output = output + output_y
|
|
@@ -424,10 +446,10 @@ class FeedForward(nn.Module):
|
|
| 424 |
dimension. Defaults to None.
|
| 425 |
|
| 426 |
Attributes:
|
| 427 |
-
w1 (
|
| 428 |
layer.
|
| 429 |
-
w2 (
|
| 430 |
-
w3 (
|
| 431 |
layer.
|
| 432 |
|
| 433 |
"""
|
|
@@ -436,21 +458,22 @@ class FeedForward(nn.Module):
|
|
| 436 |
# custom dim factor multiplier
|
| 437 |
if ffn_dim_multiplier is not None:
|
| 438 |
hidden_dim = int(ffn_dim_multiplier * hidden_dim)
|
| 439 |
-
hidden_dim = multiple_of * (
|
| 440 |
-
(hidden_dim + multiple_of - 1) // multiple_of
|
| 441 |
-
)
|
| 442 |
|
| 443 |
-
self.w1 =
|
| 444 |
-
dim,
|
| 445 |
-
|
|
|
|
| 446 |
)
|
| 447 |
-
self.w2 =
|
| 448 |
-
hidden_dim,
|
| 449 |
-
|
|
|
|
| 450 |
)
|
| 451 |
-
self.w3 =
|
| 452 |
-
dim,
|
| 453 |
-
|
|
|
|
| 454 |
)
|
| 455 |
|
| 456 |
# @torch.compile
|
|
@@ -462,9 +485,18 @@ class FeedForward(nn.Module):
|
|
| 462 |
|
| 463 |
|
| 464 |
class TransformerBlock(nn.Module):
|
| 465 |
-
def __init__(
|
| 466 |
-
|
| 467 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 468 |
"""
|
| 469 |
Initialize a TransformerBlock.
|
| 470 |
|
|
@@ -495,7 +527,9 @@ class TransformerBlock(nn.Module):
|
|
| 495 |
self.head_dim = dim // n_heads
|
| 496 |
self.attention = Attention(dim, n_heads, n_kv_heads, qk_norm, y_dim)
|
| 497 |
self.feed_forward = FeedForward(
|
| 498 |
-
dim=dim,
|
|
|
|
|
|
|
| 499 |
ffn_dim_multiplier=ffn_dim_multiplier,
|
| 500 |
)
|
| 501 |
self.layer_id = layer_id
|
|
@@ -506,9 +540,10 @@ class TransformerBlock(nn.Module):
|
|
| 506 |
|
| 507 |
self.adaLN_modulation = nn.Sequential(
|
| 508 |
nn.SiLU(),
|
| 509 |
-
|
| 510 |
-
min(dim, 1024),
|
| 511 |
-
|
|
|
|
| 512 |
),
|
| 513 |
)
|
| 514 |
|
|
@@ -536,28 +571,41 @@ class TransformerBlock(nn.Module):
|
|
| 536 |
|
| 537 |
"""
|
| 538 |
if adaln_input is not None:
|
| 539 |
-
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp =
|
| 540 |
self.adaLN_modulation(adaln_input).chunk(6, dim=1)
|
|
|
|
| 541 |
|
| 542 |
-
x = x + self.attention_norm1(
|
| 543 |
-
|
| 544 |
-
|
| 545 |
-
|
| 546 |
-
|
| 547 |
-
|
| 548 |
-
|
|
|
|
|
|
|
|
|
|
| 549 |
d = x.shape[-1]
|
| 550 |
-
x = x + self.ffn_norm1(
|
| 551 |
-
|
| 552 |
-
|
|
|
|
|
|
|
|
|
|
| 553 |
|
| 554 |
else:
|
| 555 |
-
x = x + self.attention_norm1(
|
| 556 |
-
self.
|
| 557 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 558 |
# for compatibility with torch.compile because the sequence length changes
|
| 559 |
B, L, D = x.shape
|
| 560 |
-
x = x.view(B*L, D)
|
| 561 |
x = x + self.ffn_norm1(self.feed_forward(self.ffn_norm(x)))
|
| 562 |
x = x.view(B, L, D)
|
| 563 |
|
|
@@ -566,22 +614,27 @@ class TransformerBlock(nn.Module):
|
|
| 566 |
|
| 567 |
class ParallelFinalLayer(nn.Module):
|
| 568 |
"""
|
| 569 |
-
The final layer of
|
| 570 |
"""
|
|
|
|
| 571 |
def __init__(self, hidden_size, patch_size, out_channels):
|
| 572 |
super().__init__()
|
| 573 |
self.norm_final = nn.LayerNorm(
|
| 574 |
-
hidden_size,
|
|
|
|
|
|
|
| 575 |
)
|
| 576 |
-
self.linear =
|
| 577 |
-
hidden_size,
|
| 578 |
-
|
|
|
|
| 579 |
)
|
| 580 |
self.adaLN_modulation = nn.Sequential(
|
| 581 |
nn.SiLU(),
|
| 582 |
-
|
| 583 |
-
min(hidden_size, 1024),
|
| 584 |
-
|
|
|
|
| 585 |
),
|
| 586 |
)
|
| 587 |
|
|
@@ -596,6 +649,7 @@ class NextDiT(nn.Module):
|
|
| 596 |
"""
|
| 597 |
Diffusion model with a Transformer backbone.
|
| 598 |
"""
|
|
|
|
| 599 |
def __init__(
|
| 600 |
self,
|
| 601 |
patch_size: int = 2,
|
|
@@ -610,8 +664,8 @@ class NextDiT(nn.Module):
|
|
| 610 |
learn_sigma: bool = True,
|
| 611 |
qk_norm: bool = False,
|
| 612 |
cap_feat_dim: int = 5120,
|
| 613 |
-
rope_scaling_factor: float = 1
|
| 614 |
-
ntk_factor: float=1.
|
| 615 |
) -> None:
|
| 616 |
super().__init__()
|
| 617 |
self.learn_sigma = learn_sigma
|
|
@@ -619,34 +673,49 @@ class NextDiT(nn.Module):
|
|
| 619 |
self.out_channels = in_channels * 2 if learn_sigma else in_channels
|
| 620 |
self.patch_size = patch_size
|
| 621 |
|
| 622 |
-
self.x_embedder =
|
| 623 |
in_features=patch_size * patch_size * in_channels,
|
| 624 |
out_features=dim,
|
| 625 |
bias=True,
|
| 626 |
-
gather_output=True,
|
| 627 |
-
init_method=nn.init.xavier_uniform_,
|
| 628 |
)
|
| 629 |
-
nn.init.constant_(self.x_embedder.bias, 0.)
|
| 630 |
|
| 631 |
self.t_embedder = ParallelTimestepEmbedder(min(dim, 1024))
|
| 632 |
self.cap_embedder = nn.Sequential(
|
| 633 |
nn.LayerNorm(cap_feat_dim),
|
| 634 |
-
|
| 635 |
-
|
|
|
|
|
|
|
|
|
|
| 636 |
)
|
| 637 |
|
| 638 |
-
self.layers = nn.ModuleList(
|
| 639 |
-
|
| 640 |
-
|
| 641 |
-
|
| 642 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 643 |
self.final_layer = ParallelFinalLayer(dim, patch_size, self.out_channels)
|
| 644 |
|
| 645 |
assert (dim // n_heads) % 4 == 0, "2d rope needs head dim to be divisible by 4"
|
| 646 |
self.dim = dim
|
| 647 |
self.n_heads = n_heads
|
| 648 |
self.freqs_cis = NextDiT.precompute_freqs_cis(
|
| 649 |
-
dim // n_heads,
|
|
|
|
|
|
|
|
|
|
| 650 |
)
|
| 651 |
self.rope_scaling_factor = rope_scaling_factor
|
| 652 |
self.ntk_factor = ntk_factor
|
|
@@ -655,7 +724,9 @@ class NextDiT(nn.Module):
|
|
| 655 |
# nn.init.normal_(self.eol_token, std=0.02)
|
| 656 |
nn.init.normal_(self.pad_token, std=0.02)
|
| 657 |
|
| 658 |
-
def unpatchify(
|
|
|
|
|
|
|
| 659 |
"""
|
| 660 |
x: (N, T, patch_size**2 * C)
|
| 661 |
imgs: (N, H, W, C)
|
|
@@ -673,26 +744,40 @@ class NextDiT(nn.Module):
|
|
| 673 |
for i in range(x.size(0)):
|
| 674 |
H, W = img_size[i]
|
| 675 |
L = (H // pH) * (W // pW)
|
| 676 |
-
imgs.append(
|
| 677 |
-
|
| 678 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 679 |
return imgs
|
| 680 |
|
| 681 |
def patchify_and_embed(
|
| 682 |
-
self,
|
| 683 |
-
x: List[torch.Tensor] | torch.Tensor
|
| 684 |
) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], torch.Tensor]:
|
| 685 |
self.freqs_cis = self.freqs_cis.to(x[0].device)
|
| 686 |
if isinstance(x, torch.Tensor):
|
| 687 |
pH = pW = self.patch_size
|
| 688 |
B, C, H, W = x.size()
|
| 689 |
-
x =
|
|
|
|
|
|
|
|
|
|
|
|
|
| 690 |
x = self.x_embedder(x)
|
| 691 |
x = x.flatten(1, 2)
|
| 692 |
|
| 693 |
-
mask = torch.ones(
|
|
|
|
|
|
|
| 694 |
# leave the first line for text
|
| 695 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 696 |
else:
|
| 697 |
pH = pW = self.patch_size
|
| 698 |
x_embed = []
|
|
@@ -702,30 +787,44 @@ class NextDiT(nn.Module):
|
|
| 702 |
|
| 703 |
for img in x:
|
| 704 |
C, H, W = img.size()
|
| 705 |
-
item_freqs_cis = self.freqs_cis[:H//pH, :W//pW]
|
| 706 |
-
freqs_cis.append(item_freqs_cis.flatten(0,1))
|
| 707 |
img_size.append((H, W))
|
| 708 |
-
img =
|
|
|
|
|
|
|
|
|
|
|
|
|
| 709 |
img = self.x_embedder(img)
|
| 710 |
img = img.flatten(0, 1)
|
| 711 |
l_effective_seq_len.append(len(img))
|
| 712 |
x_embed.append(img)
|
| 713 |
|
| 714 |
max_seq_len = max(l_effective_seq_len)
|
| 715 |
-
mask = torch.zeros(
|
|
|
|
|
|
|
| 716 |
padded_x_embed = []
|
| 717 |
padded_freqs_cis = []
|
| 718 |
-
for i, (item_embed, item_freqs_cis, item_seq_len) in enumerate(
|
| 719 |
-
x_embed, freqs_cis, l_effective_seq_len
|
| 720 |
-
)
|
| 721 |
-
item_embed = torch.cat(
|
| 722 |
-
|
| 723 |
-
|
| 724 |
-
|
| 725 |
-
|
| 726 |
-
|
| 727 |
-
|
| 728 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 729 |
padded_x_embed.append(item_embed)
|
| 730 |
padded_freqs_cis.append(item_freqs_cis)
|
| 731 |
mask[i][:item_seq_len] = 1
|
|
@@ -736,7 +835,7 @@ class NextDiT(nn.Module):
|
|
| 736 |
|
| 737 |
def forward(self, x, t, cap_feats, cap_mask):
|
| 738 |
"""
|
| 739 |
-
Forward pass of
|
| 740 |
t: (N,) tensor of diffusion timesteps
|
| 741 |
y: (N,) tensor of class labels
|
| 742 |
"""
|
|
@@ -746,19 +845,18 @@ class NextDiT(nn.Module):
|
|
| 746 |
|
| 747 |
# cap_freqs_cis = self.freqs_cis[:1, :cap_feats.shape[1]].to(x.device)
|
| 748 |
|
| 749 |
-
t = self.t_embedder(t)
|
| 750 |
cap_mask_float = cap_mask.float().unsqueeze(-1)
|
| 751 |
-
cap_feats_pool = (cap_feats * cap_mask_float).sum(dim=1) / cap_mask_float.sum(
|
|
|
|
|
|
|
| 752 |
cap_feats_pool = cap_feats_pool.to(cap_feats)
|
| 753 |
cap_emb = self.cap_embedder(cap_feats_pool)
|
| 754 |
adaln_input = t + cap_emb
|
| 755 |
|
| 756 |
cap_mask = cap_mask.bool()
|
| 757 |
for layer in self.layers:
|
| 758 |
-
x = layer(
|
| 759 |
-
x, mask, freqs_cis, cap_feats, cap_mask,
|
| 760 |
-
adaln_input=adaln_input
|
| 761 |
-
)
|
| 762 |
|
| 763 |
x = self.final_layer(x, adaln_input)
|
| 764 |
x = self.unpatchify(x, img_size, return_tensor=x_is_tensor)
|
|
@@ -769,25 +867,48 @@ class NextDiT(nn.Module):
|
|
| 769 |
x = [_.chunk(2, dim=0)[0] for _ in x]
|
| 770 |
return x
|
| 771 |
|
| 772 |
-
def forward_with_cfg(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 773 |
# """
|
| 774 |
-
# Forward pass of
|
| 775 |
# for classifier-free guidance.
|
| 776 |
# """
|
| 777 |
# # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
| 778 |
# print(ntk_factor, rope_scaling_factor, self.ntk_factor, self.rope_scaling_factor)
|
| 779 |
if rope_scaling_factor is not None or ntk_factor is not None:
|
| 780 |
-
rope_scaling_factor =
|
|
|
|
|
|
|
|
|
|
|
|
|
| 781 |
ntk_factor = ntk_factor if ntk_factor is not None else self.ntk_factor
|
| 782 |
-
if
|
| 783 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 784 |
self.freqs_cis = NextDiT.precompute_freqs_cis(
|
| 785 |
-
self.dim // self.n_heads,
|
| 786 |
-
|
|
|
|
|
|
|
| 787 |
)
|
| 788 |
self.rope_scaling_factor = rope_scaling_factor
|
| 789 |
self.ntk_factor = ntk_factor
|
| 790 |
-
|
| 791 |
if proportional_attn:
|
| 792 |
assert base_seqlen is not None
|
| 793 |
for layer in self.layers:
|
|
@@ -817,7 +938,7 @@ class NextDiT(nn.Module):
|
|
| 817 |
end: int,
|
| 818 |
theta: float = 10000.0,
|
| 819 |
rope_scaling_factor: float = 1.0,
|
| 820 |
-
ntk_factor: float = 1.0
|
| 821 |
):
|
| 822 |
"""
|
| 823 |
Precompute the frequency tensor for complex exponentials (cis) with
|
|
@@ -841,23 +962,27 @@ class NextDiT(nn.Module):
|
|
| 841 |
|
| 842 |
theta = theta * ntk_factor
|
| 843 |
|
| 844 |
-
logger.info(
|
| 845 |
-
|
| 846 |
-
|
| 847 |
-
|
|
|
|
|
|
|
| 848 |
t = torch.arange(end, device=freqs.device, dtype=torch.float) # type: ignore
|
| 849 |
t = t / rope_scaling_factor
|
| 850 |
freqs = torch.outer(t, freqs).float() # type: ignore
|
| 851 |
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
| 852 |
|
| 853 |
-
freqs_cis_h = freqs_cis.view(end, 1, dim//4, 1).repeat(1, end, 1, 1)
|
| 854 |
-
freqs_cis_w = freqs_cis.view(1, end, dim//4, 1).repeat(end, 1, 1, 1)
|
| 855 |
freqs_cis = torch.cat([freqs_cis_h, freqs_cis_w], dim=-1).flatten(2)
|
| 856 |
return freqs_cis
|
| 857 |
|
| 858 |
def parameter_count(self) -> int:
|
| 859 |
tensor_parallel_module_list = (
|
| 860 |
-
|
|
|
|
|
|
|
| 861 |
)
|
| 862 |
total_params = 0
|
| 863 |
|
|
@@ -865,10 +990,7 @@ class NextDiT(nn.Module):
|
|
| 865 |
nonlocal total_params
|
| 866 |
is_tp_module = isinstance(module, tensor_parallel_module_list)
|
| 867 |
for param in module.parameters(recurse=False):
|
| 868 |
-
total_params += param.numel()
|
| 869 |
-
fs_init.get_model_parallel_world_size()
|
| 870 |
-
if is_tp_module else 1
|
| 871 |
-
)
|
| 872 |
for submodule in module.children():
|
| 873 |
_recursive_count_params(submodule)
|
| 874 |
|
|
@@ -880,9 +1002,7 @@ class NextDiT(nn.Module):
|
|
| 880 |
|
| 881 |
|
| 882 |
#############################################################################
|
| 883 |
-
#
|
| 884 |
#############################################################################
|
| 885 |
def NextDiT_2B_patch2(**kwargs):
|
| 886 |
-
return NextDiT(
|
| 887 |
-
patch_size=2, dim=2304, n_layers=24, n_heads=32, **kwargs
|
| 888 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import functools
|
| 2 |
import logging
|
| 3 |
import math
|
| 4 |
from typing import Optional, Tuple, List
|
| 5 |
|
|
|
|
| 6 |
from .components import RMSNorm
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
from flash_attn import flash_attn_varlen_func
|
| 8 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
| 9 |
import torch
|
|
|
|
| 22 |
# Embedding Layers for Timesteps and Class Labels #
|
| 23 |
#############################################################################
|
| 24 |
|
| 25 |
+
|
| 26 |
class ParallelTimestepEmbedder(nn.Module):
|
| 27 |
"""
|
| 28 |
Embeds scalar timesteps into vector representations.
|
| 29 |
"""
|
| 30 |
+
|
| 31 |
def __init__(self, hidden_size, frequency_embedding_size=256):
|
| 32 |
super().__init__()
|
| 33 |
self.mlp = nn.Sequential(
|
| 34 |
+
nn.Linear(
|
| 35 |
+
frequency_embedding_size,
|
| 36 |
+
hidden_size,
|
| 37 |
+
bias=True,
|
| 38 |
),
|
| 39 |
nn.SiLU(),
|
| 40 |
+
nn.Linear(
|
| 41 |
+
hidden_size,
|
| 42 |
+
hidden_size,
|
| 43 |
+
bias=True,
|
| 44 |
),
|
| 45 |
)
|
| 46 |
self.frequency_embedding_size = frequency_embedding_size
|
|
|
|
| 58 |
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
|
| 59 |
half = dim // 2
|
| 60 |
freqs = torch.exp(
|
| 61 |
+
-math.log(max_period)
|
| 62 |
+
* torch.arange(start=0, end=half, dtype=torch.float32)
|
| 63 |
+
/ half
|
| 64 |
).to(device=t.device)
|
| 65 |
args = t[:, None].float() * freqs[None]
|
| 66 |
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
| 67 |
if dim % 2:
|
| 68 |
+
embedding = torch.cat(
|
| 69 |
+
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
|
| 70 |
+
)
|
| 71 |
return embedding
|
| 72 |
|
| 73 |
def forward(self, t):
|
|
|
|
| 80 |
r"""Embeds class labels into vector representations. Also handles label
|
| 81 |
dropout for classifier-free guidance.
|
| 82 |
"""
|
| 83 |
+
|
| 84 |
def __init__(self, num_classes, hidden_size, dropout_prob):
|
| 85 |
super().__init__()
|
| 86 |
use_cfg_embedding = int(dropout_prob > 0)
|
| 87 |
+
self.embedding_table = nn.Embedding(
|
| 88 |
+
num_classes + use_cfg_embedding,
|
| 89 |
+
hidden_size,
|
| 90 |
)
|
| 91 |
self.num_classes = num_classes
|
| 92 |
self.dropout_prob = dropout_prob
|
|
|
|
| 96 |
Drops labels to enable classifier-free guidance.
|
| 97 |
"""
|
| 98 |
if force_drop_ids is None:
|
| 99 |
+
drop_ids = (
|
| 100 |
+
torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
)
|
| 102 |
+
drop_ids = drop_ids.cuda()
|
| 103 |
drop_ids = drop_ids.to(labels.device)
|
| 104 |
else:
|
| 105 |
drop_ids = force_drop_ids == 1
|
|
|
|
| 115 |
|
| 116 |
|
| 117 |
#############################################################################
|
| 118 |
+
# Core NextDiT Model #
|
| 119 |
#############################################################################
|
| 120 |
|
| 121 |
|
| 122 |
class Attention(nn.Module):
|
| 123 |
"""Multi-head attention module."""
|
| 124 |
+
|
| 125 |
+
def __init__(
|
| 126 |
+
self,
|
| 127 |
+
dim: int,
|
| 128 |
+
n_heads: int,
|
| 129 |
+
n_kv_heads: Optional[int],
|
| 130 |
+
qk_norm: bool,
|
| 131 |
+
y_dim: int,
|
| 132 |
+
):
|
| 133 |
"""
|
| 134 |
Initialize the Attention module.
|
| 135 |
|
|
|
|
| 141 |
"""
|
| 142 |
super().__init__()
|
| 143 |
self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
|
| 144 |
+
model_parallel_size = 1
|
| 145 |
self.n_local_heads = n_heads // model_parallel_size
|
| 146 |
self.n_local_kv_heads = self.n_kv_heads // model_parallel_size
|
| 147 |
self.n_rep = self.n_local_heads // self.n_local_kv_heads
|
| 148 |
self.head_dim = dim // n_heads
|
| 149 |
|
| 150 |
+
self.wq = nn.Linear(
|
| 151 |
+
dim,
|
| 152 |
+
n_heads * self.head_dim,
|
| 153 |
+
bias=False,
|
| 154 |
)
|
| 155 |
+
self.wk = nn.Linear(
|
| 156 |
+
dim,
|
| 157 |
+
self.n_kv_heads * self.head_dim,
|
| 158 |
+
bias=False,
|
| 159 |
)
|
| 160 |
+
self.wv = nn.Linear(
|
| 161 |
+
dim,
|
| 162 |
+
self.n_kv_heads * self.head_dim,
|
| 163 |
+
bias=False,
|
| 164 |
)
|
| 165 |
if y_dim > 0:
|
| 166 |
+
self.wk_y = nn.Linear(
|
| 167 |
+
y_dim,
|
| 168 |
+
self.n_kv_heads * self.head_dim,
|
| 169 |
+
bias=False,
|
| 170 |
)
|
| 171 |
+
self.wv_y = nn.Linear(
|
| 172 |
+
y_dim,
|
| 173 |
+
self.n_kv_heads * self.head_dim,
|
| 174 |
+
bias=False,
|
| 175 |
)
|
| 176 |
self.gate = nn.Parameter(torch.zeros([self.n_local_heads]))
|
| 177 |
|
| 178 |
+
self.wo = nn.Linear(
|
| 179 |
+
n_heads * self.head_dim,
|
| 180 |
+
dim,
|
| 181 |
+
bias=False,
|
| 182 |
)
|
| 183 |
|
| 184 |
if qk_norm:
|
|
|
|
| 191 |
else:
|
| 192 |
self.q_norm = self.k_norm = nn.Identity()
|
| 193 |
self.ky_norm = nn.Identity()
|
| 194 |
+
|
| 195 |
# for proportional attention computation
|
| 196 |
self.base_seqlen = None
|
| 197 |
self.proportional_attn = False
|
|
|
|
| 221 |
ndim = x.ndim
|
| 222 |
assert 0 <= 1 < ndim
|
| 223 |
assert freqs_cis.shape == (x.shape[1], x.shape[-1])
|
| 224 |
+
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
|
|
|
|
| 225 |
return freqs_cis.view(*shape)
|
| 226 |
|
| 227 |
@staticmethod
|
|
|
|
| 255 |
return x_out.type_as(x_in)
|
| 256 |
|
| 257 |
# copied from huggingface modeling_llama.py
|
| 258 |
+
def _upad_input(
|
| 259 |
+
self, query_layer, key_layer, value_layer, attention_mask, query_length
|
| 260 |
+
):
|
| 261 |
|
| 262 |
def _get_unpad_data(attention_mask):
|
| 263 |
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
| 264 |
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
| 265 |
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
| 266 |
+
cu_seqlens = F.pad(
|
| 267 |
+
torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)
|
| 268 |
+
)
|
| 269 |
return (
|
| 270 |
indices,
|
| 271 |
cu_seqlens,
|
|
|
|
| 276 |
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
| 277 |
|
| 278 |
key_layer = index_first_axis(
|
| 279 |
+
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
|
| 280 |
+
indices_k,
|
| 281 |
)
|
| 282 |
value_layer = index_first_axis(
|
| 283 |
+
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
|
| 284 |
+
indices_k,
|
| 285 |
)
|
| 286 |
if query_length == kv_seq_len:
|
| 287 |
query_layer = index_first_axis(
|
| 288 |
+
query_layer.reshape(
|
| 289 |
+
batch_size * kv_seq_len, self.n_local_heads, head_dim
|
| 290 |
+
),
|
| 291 |
+
indices_k,
|
| 292 |
)
|
| 293 |
cu_seqlens_q = cu_seqlens_k
|
| 294 |
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
|
|
|
| 303 |
else:
|
| 304 |
# The -q_len: slice assumes left padding.
|
| 305 |
attention_mask = attention_mask[:, -query_length:]
|
| 306 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(
|
| 307 |
+
query_layer, attention_mask
|
| 308 |
+
)
|
| 309 |
|
| 310 |
return (
|
| 311 |
query_layer,
|
|
|
|
| 354 |
|
| 355 |
if dtype in [torch.float16, torch.bfloat16]:
|
| 356 |
# begin var_len flash attn
|
| 357 |
+
(
|
| 358 |
+
query_states,
|
| 359 |
+
key_states,
|
| 360 |
+
value_states,
|
| 361 |
+
indices_q,
|
| 362 |
+
cu_seq_lens,
|
| 363 |
+
max_seq_lens,
|
| 364 |
+
) = self._upad_input(xq, xk, xv, x_mask, seqlen)
|
| 365 |
|
| 366 |
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
| 367 |
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
| 368 |
|
| 369 |
if self.proportional_attn:
|
| 370 |
+
softmax_scale = math.sqrt(
|
| 371 |
+
math.log(seqlen, self.base_seqlen) / self.head_dim
|
| 372 |
+
)
|
| 373 |
else:
|
| 374 |
softmax_scale = math.sqrt(1 / self.head_dim)
|
| 375 |
|
|
|
|
| 381 |
cu_seqlens_k=cu_seqlens_k,
|
| 382 |
max_seqlen_q=max_seqlen_in_batch_q,
|
| 383 |
max_seqlen_k=max_seqlen_in_batch_k,
|
| 384 |
+
dropout_p=0.0,
|
| 385 |
causal=False,
|
| 386 |
+
softmax_scale=softmax_scale,
|
| 387 |
)
|
| 388 |
output = pad_input(attn_output_unpad, indices_q, bsz, seqlen)
|
| 389 |
# end var_len_flash_attn
|
| 390 |
|
| 391 |
else:
|
| 392 |
+
output = (
|
| 393 |
+
F.scaled_dot_product_attention(
|
| 394 |
+
xq.permute(0, 2, 1, 3),
|
| 395 |
+
xk.permute(0, 2, 1, 3),
|
| 396 |
+
xv.permute(0, 2, 1, 3),
|
| 397 |
+
attn_mask=x_mask.bool()
|
| 398 |
+
.view(bsz, 1, 1, seqlen)
|
| 399 |
+
.expand(-1, self.n_local_heads, seqlen, -1),
|
| 400 |
+
)
|
| 401 |
+
.permute(0, 2, 1, 3)
|
| 402 |
+
.to(dtype)
|
| 403 |
+
)
|
| 404 |
|
| 405 |
if hasattr(self, "wk_y"):
|
| 406 |
# todo better flash_attn support
|
| 407 |
+
yk = self.ky_norm(self.wk_y(y)).view(
|
| 408 |
+
bsz, -1, self.n_local_kv_heads, self.head_dim
|
| 409 |
+
)
|
| 410 |
yv = self.wv_y(y).view(bsz, -1, self.n_local_kv_heads, self.head_dim)
|
| 411 |
n_rep = self.n_local_heads // self.n_local_kv_heads
|
| 412 |
if n_rep >= 1:
|
|
|
|
| 416 |
xq.permute(0, 2, 1, 3),
|
| 417 |
yk.permute(0, 2, 1, 3),
|
| 418 |
yv.permute(0, 2, 1, 3),
|
| 419 |
+
y_mask.view(bsz, 1, 1, -1).expand(bsz, self.n_local_heads, seqlen, -1),
|
| 420 |
).permute(0, 2, 1, 3)
|
| 421 |
output_y = output_y * self.gate.tanh().view(1, 1, -1, 1)
|
| 422 |
output = output + output_y
|
|
|
|
| 446 |
dimension. Defaults to None.
|
| 447 |
|
| 448 |
Attributes:
|
| 449 |
+
w1 (nn.Linear): Linear transformation for the first
|
| 450 |
layer.
|
| 451 |
+
w2 (nn.Linear): Linear transformation for the second layer.
|
| 452 |
+
w3 (nn.Linear): Linear transformation for the third
|
| 453 |
layer.
|
| 454 |
|
| 455 |
"""
|
|
|
|
| 458 |
# custom dim factor multiplier
|
| 459 |
if ffn_dim_multiplier is not None:
|
| 460 |
hidden_dim = int(ffn_dim_multiplier * hidden_dim)
|
| 461 |
+
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
|
|
|
|
|
|
|
| 462 |
|
| 463 |
+
self.w1 = nn.Linear(
|
| 464 |
+
dim,
|
| 465 |
+
hidden_dim,
|
| 466 |
+
bias=False,
|
| 467 |
)
|
| 468 |
+
self.w2 = nn.Linear(
|
| 469 |
+
hidden_dim,
|
| 470 |
+
dim,
|
| 471 |
+
bias=False,
|
| 472 |
)
|
| 473 |
+
self.w3 = nn.Linear(
|
| 474 |
+
dim,
|
| 475 |
+
hidden_dim,
|
| 476 |
+
bias=False,
|
| 477 |
)
|
| 478 |
|
| 479 |
# @torch.compile
|
|
|
|
| 485 |
|
| 486 |
|
| 487 |
class TransformerBlock(nn.Module):
|
| 488 |
+
def __init__(
|
| 489 |
+
self,
|
| 490 |
+
layer_id: int,
|
| 491 |
+
dim: int,
|
| 492 |
+
n_heads: int,
|
| 493 |
+
n_kv_heads: int,
|
| 494 |
+
multiple_of: int,
|
| 495 |
+
ffn_dim_multiplier: float,
|
| 496 |
+
norm_eps: float,
|
| 497 |
+
qk_norm: bool,
|
| 498 |
+
y_dim: int,
|
| 499 |
+
) -> None:
|
| 500 |
"""
|
| 501 |
Initialize a TransformerBlock.
|
| 502 |
|
|
|
|
| 527 |
self.head_dim = dim // n_heads
|
| 528 |
self.attention = Attention(dim, n_heads, n_kv_heads, qk_norm, y_dim)
|
| 529 |
self.feed_forward = FeedForward(
|
| 530 |
+
dim=dim,
|
| 531 |
+
hidden_dim=4 * dim,
|
| 532 |
+
multiple_of=multiple_of,
|
| 533 |
ffn_dim_multiplier=ffn_dim_multiplier,
|
| 534 |
)
|
| 535 |
self.layer_id = layer_id
|
|
|
|
| 540 |
|
| 541 |
self.adaLN_modulation = nn.Sequential(
|
| 542 |
nn.SiLU(),
|
| 543 |
+
nn.Linear(
|
| 544 |
+
min(dim, 1024),
|
| 545 |
+
6 * dim,
|
| 546 |
+
bias=True,
|
| 547 |
),
|
| 548 |
)
|
| 549 |
|
|
|
|
| 571 |
|
| 572 |
"""
|
| 573 |
if adaln_input is not None:
|
| 574 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
|
| 575 |
self.adaLN_modulation(adaln_input).chunk(6, dim=1)
|
| 576 |
+
)
|
| 577 |
|
| 578 |
+
x = x + self.attention_norm1(
|
| 579 |
+
gate_msa.unsqueeze(1)
|
| 580 |
+
* self.attention(
|
| 581 |
+
modulate(self.attention_norm(x), shift_msa, scale_msa),
|
| 582 |
+
x_mask,
|
| 583 |
+
freqs_cis,
|
| 584 |
+
self.attention_y_norm(y),
|
| 585 |
+
y_mask,
|
| 586 |
+
)
|
| 587 |
+
)
|
| 588 |
d = x.shape[-1]
|
| 589 |
+
x = x + self.ffn_norm1(
|
| 590 |
+
gate_mlp.unsqueeze(1)
|
| 591 |
+
* self.feed_forward(
|
| 592 |
+
modulate(self.ffn_norm(x), shift_mlp, scale_mlp).view(-1, d),
|
| 593 |
+
).view(*x.shape)
|
| 594 |
+
)
|
| 595 |
|
| 596 |
else:
|
| 597 |
+
x = x + self.attention_norm1(
|
| 598 |
+
self.attention(
|
| 599 |
+
self.attention_norm(x),
|
| 600 |
+
x_mask,
|
| 601 |
+
freqs_cis,
|
| 602 |
+
self.attention_y_norm(y),
|
| 603 |
+
y_mask,
|
| 604 |
+
)
|
| 605 |
+
)
|
| 606 |
# for compatibility with torch.compile because the sequence length changes
|
| 607 |
B, L, D = x.shape
|
| 608 |
+
x = x.view(B * L, D)
|
| 609 |
x = x + self.ffn_norm1(self.feed_forward(self.ffn_norm(x)))
|
| 610 |
x = x.view(B, L, D)
|
| 611 |
|
|
|
|
| 614 |
|
| 615 |
class ParallelFinalLayer(nn.Module):
|
| 616 |
"""
|
| 617 |
+
The final layer of NextDiT.
|
| 618 |
"""
|
| 619 |
+
|
| 620 |
def __init__(self, hidden_size, patch_size, out_channels):
|
| 621 |
super().__init__()
|
| 622 |
self.norm_final = nn.LayerNorm(
|
| 623 |
+
hidden_size,
|
| 624 |
+
elementwise_affine=False,
|
| 625 |
+
eps=1e-6,
|
| 626 |
)
|
| 627 |
+
self.linear = nn.Linear(
|
| 628 |
+
hidden_size,
|
| 629 |
+
patch_size * patch_size * out_channels,
|
| 630 |
+
bias=True,
|
| 631 |
)
|
| 632 |
self.adaLN_modulation = nn.Sequential(
|
| 633 |
nn.SiLU(),
|
| 634 |
+
nn.Linear(
|
| 635 |
+
min(hidden_size, 1024),
|
| 636 |
+
2 * hidden_size,
|
| 637 |
+
bias=True,
|
| 638 |
),
|
| 639 |
)
|
| 640 |
|
|
|
|
| 649 |
"""
|
| 650 |
Diffusion model with a Transformer backbone.
|
| 651 |
"""
|
| 652 |
+
|
| 653 |
def __init__(
|
| 654 |
self,
|
| 655 |
patch_size: int = 2,
|
|
|
|
| 664 |
learn_sigma: bool = True,
|
| 665 |
qk_norm: bool = False,
|
| 666 |
cap_feat_dim: int = 5120,
|
| 667 |
+
rope_scaling_factor: float = 1.0,
|
| 668 |
+
ntk_factor: float = 1.0,
|
| 669 |
) -> None:
|
| 670 |
super().__init__()
|
| 671 |
self.learn_sigma = learn_sigma
|
|
|
|
| 673 |
self.out_channels = in_channels * 2 if learn_sigma else in_channels
|
| 674 |
self.patch_size = patch_size
|
| 675 |
|
| 676 |
+
self.x_embedder = nn.Linear(
|
| 677 |
in_features=patch_size * patch_size * in_channels,
|
| 678 |
out_features=dim,
|
| 679 |
bias=True,
|
|
|
|
|
|
|
| 680 |
)
|
| 681 |
+
nn.init.constant_(self.x_embedder.bias, 0.0)
|
| 682 |
|
| 683 |
self.t_embedder = ParallelTimestepEmbedder(min(dim, 1024))
|
| 684 |
self.cap_embedder = nn.Sequential(
|
| 685 |
nn.LayerNorm(cap_feat_dim),
|
| 686 |
+
nn.Linear(
|
| 687 |
+
cap_feat_dim,
|
| 688 |
+
min(dim, 1024),
|
| 689 |
+
bias=True,
|
| 690 |
+
),
|
| 691 |
)
|
| 692 |
|
| 693 |
+
self.layers = nn.ModuleList(
|
| 694 |
+
[
|
| 695 |
+
TransformerBlock(
|
| 696 |
+
layer_id,
|
| 697 |
+
dim,
|
| 698 |
+
n_heads,
|
| 699 |
+
n_kv_heads,
|
| 700 |
+
multiple_of,
|
| 701 |
+
ffn_dim_multiplier,
|
| 702 |
+
norm_eps,
|
| 703 |
+
qk_norm,
|
| 704 |
+
cap_feat_dim,
|
| 705 |
+
)
|
| 706 |
+
for layer_id in range(n_layers)
|
| 707 |
+
]
|
| 708 |
+
)
|
| 709 |
self.final_layer = ParallelFinalLayer(dim, patch_size, self.out_channels)
|
| 710 |
|
| 711 |
assert (dim // n_heads) % 4 == 0, "2d rope needs head dim to be divisible by 4"
|
| 712 |
self.dim = dim
|
| 713 |
self.n_heads = n_heads
|
| 714 |
self.freqs_cis = NextDiT.precompute_freqs_cis(
|
| 715 |
+
dim // n_heads,
|
| 716 |
+
384,
|
| 717 |
+
rope_scaling_factor=rope_scaling_factor,
|
| 718 |
+
ntk_factor=ntk_factor,
|
| 719 |
)
|
| 720 |
self.rope_scaling_factor = rope_scaling_factor
|
| 721 |
self.ntk_factor = ntk_factor
|
|
|
|
| 724 |
# nn.init.normal_(self.eol_token, std=0.02)
|
| 725 |
nn.init.normal_(self.pad_token, std=0.02)
|
| 726 |
|
| 727 |
+
def unpatchify(
|
| 728 |
+
self, x: torch.Tensor, img_size: List[Tuple[int, int]], return_tensor=False
|
| 729 |
+
) -> List[torch.Tensor]:
|
| 730 |
"""
|
| 731 |
x: (N, T, patch_size**2 * C)
|
| 732 |
imgs: (N, H, W, C)
|
|
|
|
| 744 |
for i in range(x.size(0)):
|
| 745 |
H, W = img_size[i]
|
| 746 |
L = (H // pH) * (W // pW)
|
| 747 |
+
imgs.append(
|
| 748 |
+
x[i][:L]
|
| 749 |
+
.view(H // pH, W // pW, pH, pW, self.out_channels)
|
| 750 |
+
.permute(4, 0, 2, 1, 3)
|
| 751 |
+
.flatten(3, 4)
|
| 752 |
+
.flatten(1, 2)
|
| 753 |
+
)
|
| 754 |
return imgs
|
| 755 |
|
| 756 |
def patchify_and_embed(
|
| 757 |
+
self, x: List[torch.Tensor] | torch.Tensor
|
|
|
|
| 758 |
) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], torch.Tensor]:
|
| 759 |
self.freqs_cis = self.freqs_cis.to(x[0].device)
|
| 760 |
if isinstance(x, torch.Tensor):
|
| 761 |
pH = pW = self.patch_size
|
| 762 |
B, C, H, W = x.size()
|
| 763 |
+
x = (
|
| 764 |
+
x.view(B, C, H // pH, pH, W // pW, pW)
|
| 765 |
+
.permute(0, 2, 4, 1, 3, 5)
|
| 766 |
+
.flatten(3)
|
| 767 |
+
)
|
| 768 |
x = self.x_embedder(x)
|
| 769 |
x = x.flatten(1, 2)
|
| 770 |
|
| 771 |
+
mask = torch.ones(
|
| 772 |
+
x.shape[0], x.shape[1], dtype=torch.int32, device=x.device
|
| 773 |
+
)
|
| 774 |
# leave the first line for text
|
| 775 |
+
return (
|
| 776 |
+
x,
|
| 777 |
+
mask,
|
| 778 |
+
[(H, W)] * B,
|
| 779 |
+
self.freqs_cis[: H // pH, : W // pW].flatten(0, 1).unsqueeze(0),
|
| 780 |
+
)
|
| 781 |
else:
|
| 782 |
pH = pW = self.patch_size
|
| 783 |
x_embed = []
|
|
|
|
| 787 |
|
| 788 |
for img in x:
|
| 789 |
C, H, W = img.size()
|
| 790 |
+
item_freqs_cis = self.freqs_cis[: H // pH, : W // pW]
|
| 791 |
+
freqs_cis.append(item_freqs_cis.flatten(0, 1))
|
| 792 |
img_size.append((H, W))
|
| 793 |
+
img = (
|
| 794 |
+
img.view(C, H // pH, pH, W // pW, pW)
|
| 795 |
+
.permute(1, 3, 0, 2, 4)
|
| 796 |
+
.flatten(2)
|
| 797 |
+
)
|
| 798 |
img = self.x_embedder(img)
|
| 799 |
img = img.flatten(0, 1)
|
| 800 |
l_effective_seq_len.append(len(img))
|
| 801 |
x_embed.append(img)
|
| 802 |
|
| 803 |
max_seq_len = max(l_effective_seq_len)
|
| 804 |
+
mask = torch.zeros(
|
| 805 |
+
len(x), max_seq_len, dtype=torch.int32, device=x[0].device
|
| 806 |
+
)
|
| 807 |
padded_x_embed = []
|
| 808 |
padded_freqs_cis = []
|
| 809 |
+
for i, (item_embed, item_freqs_cis, item_seq_len) in enumerate(
|
| 810 |
+
zip(x_embed, freqs_cis, l_effective_seq_len)
|
| 811 |
+
):
|
| 812 |
+
item_embed = torch.cat(
|
| 813 |
+
[
|
| 814 |
+
item_embed,
|
| 815 |
+
self.pad_token.view(1, -1).expand(
|
| 816 |
+
max_seq_len - item_seq_len, -1
|
| 817 |
+
),
|
| 818 |
+
],
|
| 819 |
+
dim=0,
|
| 820 |
+
)
|
| 821 |
+
item_freqs_cis = torch.cat(
|
| 822 |
+
[
|
| 823 |
+
item_freqs_cis,
|
| 824 |
+
item_freqs_cis[-1:].expand(max_seq_len - item_seq_len, -1),
|
| 825 |
+
],
|
| 826 |
+
dim=0,
|
| 827 |
+
)
|
| 828 |
padded_x_embed.append(item_embed)
|
| 829 |
padded_freqs_cis.append(item_freqs_cis)
|
| 830 |
mask[i][:item_seq_len] = 1
|
|
|
|
| 835 |
|
| 836 |
def forward(self, x, t, cap_feats, cap_mask):
|
| 837 |
"""
|
| 838 |
+
Forward pass of NextDiT.
|
| 839 |
t: (N,) tensor of diffusion timesteps
|
| 840 |
y: (N,) tensor of class labels
|
| 841 |
"""
|
|
|
|
| 845 |
|
| 846 |
# cap_freqs_cis = self.freqs_cis[:1, :cap_feats.shape[1]].to(x.device)
|
| 847 |
|
| 848 |
+
t = self.t_embedder(t) # (N, D)
|
| 849 |
cap_mask_float = cap_mask.float().unsqueeze(-1)
|
| 850 |
+
cap_feats_pool = (cap_feats * cap_mask_float).sum(dim=1) / cap_mask_float.sum(
|
| 851 |
+
dim=1
|
| 852 |
+
)
|
| 853 |
cap_feats_pool = cap_feats_pool.to(cap_feats)
|
| 854 |
cap_emb = self.cap_embedder(cap_feats_pool)
|
| 855 |
adaln_input = t + cap_emb
|
| 856 |
|
| 857 |
cap_mask = cap_mask.bool()
|
| 858 |
for layer in self.layers:
|
| 859 |
+
x = layer(x, mask, freqs_cis, cap_feats, cap_mask, adaln_input=adaln_input)
|
|
|
|
|
|
|
|
|
|
| 860 |
|
| 861 |
x = self.final_layer(x, adaln_input)
|
| 862 |
x = self.unpatchify(x, img_size, return_tensor=x_is_tensor)
|
|
|
|
| 867 |
x = [_.chunk(2, dim=0)[0] for _ in x]
|
| 868 |
return x
|
| 869 |
|
| 870 |
+
def forward_with_cfg(
|
| 871 |
+
self,
|
| 872 |
+
x,
|
| 873 |
+
t,
|
| 874 |
+
cap_feats,
|
| 875 |
+
cap_mask,
|
| 876 |
+
cfg_scale,
|
| 877 |
+
rope_scaling_factor=None,
|
| 878 |
+
ntk_factor=None,
|
| 879 |
+
base_seqlen: Optional[int] = None,
|
| 880 |
+
proportional_attn: bool = False,
|
| 881 |
+
):
|
| 882 |
# """
|
| 883 |
+
# Forward pass of NextDiT, but also batches the unconNextditional forward pass
|
| 884 |
# for classifier-free guidance.
|
| 885 |
# """
|
| 886 |
# # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
|
| 887 |
# print(ntk_factor, rope_scaling_factor, self.ntk_factor, self.rope_scaling_factor)
|
| 888 |
if rope_scaling_factor is not None or ntk_factor is not None:
|
| 889 |
+
rope_scaling_factor = (
|
| 890 |
+
rope_scaling_factor
|
| 891 |
+
if rope_scaling_factor is not None
|
| 892 |
+
else self.rope_scaling_factor
|
| 893 |
+
)
|
| 894 |
ntk_factor = ntk_factor if ntk_factor is not None else self.ntk_factor
|
| 895 |
+
if (
|
| 896 |
+
rope_scaling_factor != self.rope_scaling_factor
|
| 897 |
+
or ntk_factor != self.ntk_factor
|
| 898 |
+
):
|
| 899 |
+
print(
|
| 900 |
+
f"override freqs_cis, rope_scaling {rope_scaling_factor}, ntk {ntk_factor}",
|
| 901 |
+
flush=True,
|
| 902 |
+
)
|
| 903 |
self.freqs_cis = NextDiT.precompute_freqs_cis(
|
| 904 |
+
self.dim // self.n_heads,
|
| 905 |
+
384,
|
| 906 |
+
rope_scaling_factor=rope_scaling_factor,
|
| 907 |
+
ntk_factor=ntk_factor,
|
| 908 |
)
|
| 909 |
self.rope_scaling_factor = rope_scaling_factor
|
| 910 |
self.ntk_factor = ntk_factor
|
| 911 |
+
|
| 912 |
if proportional_attn:
|
| 913 |
assert base_seqlen is not None
|
| 914 |
for layer in self.layers:
|
|
|
|
| 938 |
end: int,
|
| 939 |
theta: float = 10000.0,
|
| 940 |
rope_scaling_factor: float = 1.0,
|
| 941 |
+
ntk_factor: float = 1.0,
|
| 942 |
):
|
| 943 |
"""
|
| 944 |
Precompute the frequency tensor for complex exponentials (cis) with
|
|
|
|
| 962 |
|
| 963 |
theta = theta * ntk_factor
|
| 964 |
|
| 965 |
+
logger.info(
|
| 966 |
+
f"theta {theta} rope scaling {rope_scaling_factor} ntk {ntk_factor}"
|
| 967 |
+
)
|
| 968 |
+
freqs = 1.0 / (
|
| 969 |
+
theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float().cuda() / dim)
|
| 970 |
+
)
|
| 971 |
t = torch.arange(end, device=freqs.device, dtype=torch.float) # type: ignore
|
| 972 |
t = t / rope_scaling_factor
|
| 973 |
freqs = torch.outer(t, freqs).float() # type: ignore
|
| 974 |
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
| 975 |
|
| 976 |
+
freqs_cis_h = freqs_cis.view(end, 1, dim // 4, 1).repeat(1, end, 1, 1)
|
| 977 |
+
freqs_cis_w = freqs_cis.view(1, end, dim // 4, 1).repeat(end, 1, 1, 1)
|
| 978 |
freqs_cis = torch.cat([freqs_cis_h, freqs_cis_w], dim=-1).flatten(2)
|
| 979 |
return freqs_cis
|
| 980 |
|
| 981 |
def parameter_count(self) -> int:
|
| 982 |
tensor_parallel_module_list = (
|
| 983 |
+
nn.Linear,
|
| 984 |
+
nn.Linear,
|
| 985 |
+
nn.Embedding,
|
| 986 |
)
|
| 987 |
total_params = 0
|
| 988 |
|
|
|
|
| 990 |
nonlocal total_params
|
| 991 |
is_tp_module = isinstance(module, tensor_parallel_module_list)
|
| 992 |
for param in module.parameters(recurse=False):
|
| 993 |
+
total_params += param.numel()
|
|
|
|
|
|
|
|
|
|
| 994 |
for submodule in module.children():
|
| 995 |
_recursive_count_params(submodule)
|
| 996 |
|
|
|
|
| 1002 |
|
| 1003 |
|
| 1004 |
#############################################################################
|
| 1005 |
+
# NextDiT Configs #
|
| 1006 |
#############################################################################
|
| 1007 |
def NextDiT_2B_patch2(**kwargs):
|
| 1008 |
+
return NextDiT(patch_size=2, dim=2304, n_layers=24, n_heads=32, **kwargs)
|
|
|
|
|
|