File size: 8,195 Bytes
e181cae | 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 | # -*- coding: utf-8 -*-
import torch
import torch.cuda.amp as amp
import torch.nn as nn
from diffusers.configuration_utils import register_to_config
from wan.modules.model import WanModel, WanAttentionBlock, sinusoidal_embedding_1d
import torch.utils.checkpoint
def create_custom_forward(module):
def custom_forward(*inputs, **kwargs):
return module(*inputs, **kwargs)
return custom_forward
def gradient_checkpoint_forward(
model,
use_gradient_checkpointing,
use_gradient_checkpointing_offload,
*args,
**kwargs,
):
if use_gradient_checkpointing_offload:
with torch.autograd.graph.save_on_cpu():
model_output = torch.utils.checkpoint.checkpoint(
create_custom_forward(model),
*args,
**kwargs,
use_reentrant=False,
)
elif use_gradient_checkpointing:
model_output = torch.utils.checkpoint.checkpoint(
create_custom_forward(model),
*args,
**kwargs,
use_reentrant=False,
)
else:
model_output = model(*args, **kwargs)
return model_output
class VaceWanAttentionBlock(WanAttentionBlock):
def __init__(
self,
cross_attn_type,
dim,
ffn_dim,
num_heads,
window_size=(-1, -1),
qk_norm=True,
cross_attn_norm=False,
eps=1e-6,
block_id=0
):
super().__init__(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps)
self.block_id = block_id
if block_id == 0:
self.before_proj = nn.Linear(self.dim, self.dim)
nn.init.zeros_(self.before_proj.weight)
nn.init.zeros_(self.before_proj.bias)
self.after_proj = nn.Linear(self.dim, self.dim)
nn.init.zeros_(self.after_proj.weight)
nn.init.zeros_(self.after_proj.bias)
def forward(self, c, x, **kwargs):
if self.block_id == 0:
c = self.before_proj(c) + x
all_c = []
else:
all_c = list(torch.unbind(c))
c = all_c.pop(-1)
c = super().forward(c, **kwargs)
c_skip = self.after_proj(c)
all_c += [c_skip, c]
c = torch.stack(all_c)
return c
class BaseWanAttentionBlock(WanAttentionBlock):
def __init__(
self,
cross_attn_type,
dim,
ffn_dim,
num_heads,
window_size=(-1, -1),
qk_norm=True,
cross_attn_norm=False,
eps=1e-6,
block_id=None
):
super().__init__(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps)
self.block_id = block_id
def forward(self, x, hints, context_scale=1.0, **kwargs):
x = super().forward(x, **kwargs)
if self.block_id is not None:
x = x + hints[self.block_id] * context_scale
return x
class VaceWanModel(WanModel):
@register_to_config
def __init__(self,
vace_layers=None,
vace_in_dim=None,
model_type='t2v',
patch_size=(1, 2, 2),
text_len=512,
in_dim=16,
dim=2048,
ffn_dim=8192,
freq_dim=256,
text_dim=4096,
out_dim=16,
num_heads=16,
num_layers=32,
window_size=(-1, -1),
qk_norm=True,
cross_attn_norm=True,
eps=1e-6):
model_type = "t2v" # TODO: Hard code for both preview and official versions.
super().__init__(model_type, patch_size, text_len, in_dim, dim, ffn_dim, freq_dim, text_dim, out_dim,
num_heads, num_layers, window_size, qk_norm, cross_attn_norm, eps)
self.vace_layers = [i for i in range(0, self.num_layers, 2)] if vace_layers is None else vace_layers
self.vace_in_dim = self.in_dim if vace_in_dim is None else vace_in_dim
assert 0 in self.vace_layers
self.vace_layers_mapping = {i: n for n, i in enumerate(self.vace_layers)}
# blocks
self.blocks = nn.ModuleList([
BaseWanAttentionBlock('t2v_cross_attn', self.dim, self.ffn_dim, self.num_heads, self.window_size, self.qk_norm,
self.cross_attn_norm, self.eps,
block_id=self.vace_layers_mapping[i] if i in self.vace_layers else None)
for i in range(self.num_layers)
])
# vace blocks
self.vace_blocks = nn.ModuleList([
VaceWanAttentionBlock('t2v_cross_attn', self.dim, self.ffn_dim, self.num_heads, self.window_size, self.qk_norm,
self.cross_attn_norm, self.eps, block_id=i)
for i in self.vace_layers
])
# vace patch embeddings
self.vace_patch_embedding = nn.Conv3d(
self.vace_in_dim, self.dim, kernel_size=self.patch_size, stride=self.patch_size
)
def forward_vace(
self,
x,
vace_context,
seq_len,
kwargs,
use_gradient_checkpointing=False,
use_gradient_checkpointing_offload=False,
):
# embeddings
c = [self.vace_patch_embedding(u.unsqueeze(0)) for u in vace_context]
c = [u.flatten(2).transpose(1, 2) for u in c]
c = torch.cat([
torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))],
dim=1) for u in c
])
new_kwargs = dict(x=x)
new_kwargs.update(kwargs)
for block in self.vace_blocks:
c = gradient_checkpoint_forward(
block,
use_gradient_checkpointing,
use_gradient_checkpointing_offload,
c,
**new_kwargs,
)
hints = torch.unbind(c)[:-1]
return hints
def forward(
self,
x,
t,
vace_context,
context,
seq_len,
vace_context_scale=1.0,
clip_fea=None,
y=None,
use_gradient_checkpointing=True,
use_gradient_checkpointing_offload=False,
):
device = self.patch_embedding.weight.device
if self.freqs.device != device:
self.freqs = self.freqs.to(device)
# embeddings
x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
grid_sizes = torch.stack(
[torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
x = [u.flatten(2).transpose(1, 2) for u in x]
seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
assert seq_lens.max() <= seq_len
x = torch.cat([
torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))],
dim=1) for u in x
])
# time embeddings
with amp.autocast(dtype=torch.float32):
e = self.time_embedding(
sinusoidal_embedding_1d(self.freq_dim, t).float())
e0 = self.time_projection(e).unflatten(1, (6, self.dim))
# text embedding
context_lens = None
context = self.text_embedding(context)
kwargs = dict(
e=e0,
seq_lens=seq_lens,
grid_sizes=grid_sizes,
freqs=self.freqs,
context=context,
context_lens=context_lens,
)
hints = self.forward_vace(
x, vace_context, seq_len, kwargs,
use_gradient_checkpointing=False,
use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
)
kwargs['hints'] = hints
kwargs['context_scale'] = vace_context_scale
for block in self.blocks:
x = gradient_checkpoint_forward(
block,
use_gradient_checkpointing,
use_gradient_checkpointing_offload,
x,
**kwargs,
)
x = self.head(x, e)
x = self.unpatchify(x, grid_sizes)
return [u.float() for u in x]
|