File size: 10,023 Bytes
70ceb8b | 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 | from typing import Dict, Union
import torch
import torch.nn as nn
from ..wanvideo.modules.model import WanLayerNorm, WanSelfAttention, EmbedND_RifleX, sinusoidal_embedding_1d, apply_rotary_emb_split, apply_rope_comfy1
class WanAttentionBlock(nn.Module):
def __init__(self, in_features, out_features, ffn_dim, ffn2_dim, num_heads, qk_norm=True, cross_attn_norm=False, eps=1e-6, attention_mode="sdpa", rope_func="comfy", rms_norm_function="default"):
super().__init__()
self.dim = out_features
self.ffn_dim = ffn_dim
self.num_heads = num_heads
self.head_dim = out_features // num_heads
self.qk_norm = qk_norm
self.cross_attn_norm = cross_attn_norm
self.eps = eps
self.attention_mode = attention_mode
self.rope_func = rope_func
# layers
self.norm1 = WanLayerNorm(self.dim, eps)
self.self_attn = WanSelfAttention(in_features, out_features, num_heads, qk_norm, eps, self.attention_mode, rms_norm_function=rms_norm_function, head_norm=False)
self.norm2 = WanLayerNorm(self.dim, eps)
self.ffn = nn.Sequential(nn.Linear(in_features, ffn_dim), nn.GELU(approximate='tanh'), nn.Linear(ffn2_dim, out_features))
self.modulation = nn.Parameter(torch.randn(1, 6, out_features) / in_features**0.5)
def get_mod(self, e, modulation):
if e.dim() == 3:
if e.shape[-1] == 512:
e = self.modulation(e)
return e.unsqueeze(2).chunk(6, dim=-1)
return (modulation + e).chunk(6, dim=1) # 1, 6, dim
elif e.dim() == 4:
e_mod = modulation.unsqueeze(2) + e
return [ei.squeeze(1) for ei in e_mod.unbind(dim=1)]
def modulate(self, norm_x, shift_msa, scale_msa):
return torch.addcmul(shift_msa, norm_x, 1 + scale_msa)
def ffn_chunked(self, mod_x, num_chunks=4):
seq_len = mod_x.shape[1]
if seq_len <= 8192 or num_chunks <= 1:
return self.ffn(mod_x)
return torch.cat([self.ffn(chunk.contiguous()) for chunk in mod_x.chunk(num_chunks, dim=1)], dim=1)
#region attention forward
def forward(self, x, e, seq_lens, freqs, split_rope=True, e_tr=None, tr_start=0, tr_num=0):
use_token_replace = False
if e_tr is not None and tr_num > 0:
tr_shift_msa, tr_scale_msa, tr_gate_msa, tr_shift_mlp, tr_scale_mlp, tr_gate_mlp = self.get_mod(e_tr.to(x.device), self.modulation)
use_token_replace = True
tr_start = tr_start or 0
tr_end = tr_start + (tr_num or 0)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.get_mod(e.to(x.device), self.modulation)
del e
input_dtype = x.dtype
if use_token_replace:
norm_x = self.norm1(x.to(shift_msa.dtype))
input_x = torch.cat([
torch.addcmul(shift_msa, norm_x[:, :tr_start], 1 + scale_msa), # before replace → T
torch.addcmul(tr_shift_msa, norm_x[:, tr_start:tr_end], 1 + tr_scale_msa), # replace segment → t=0
torch.addcmul(shift_msa, norm_x[:, tr_end:], 1 + scale_msa) # after replace → T
], dim=1).to(input_dtype)
else:
input_x = self.modulate(self.norm1(x.to(shift_msa.dtype)), shift_msa, scale_msa).to(input_dtype)
del shift_msa, scale_msa
b, s, n, d = *x.shape[:2], self.self_attn.num_heads, self.self_attn.head_dim
h_dim = w_dim = 2 * (self.head_dim // 6)
t_dim = self.head_dim - h_dim - w_dim
q = self.self_attn.norm_q(self.self_attn.q(input_x)).to(self.self_attn.norm_q.weight.dtype).view(b, s, n, d)
if split_rope:
q = apply_rotary_emb_split(q, freqs, t_dim) # Apply split rotary embedding (only to H/W dimensions, leaving T unchanged)
else:
q = apply_rope_comfy1(q, freqs)
k = self.self_attn.norm_k(self.self_attn.k(input_x).to(self.self_attn.norm_k.weight.dtype)).to(input_x.dtype).view(b, s, n, d)
if split_rope:
k = apply_rotary_emb_split(k, freqs, t_dim)
else:
k = apply_rope_comfy1(k, freqs)
v = self.self_attn.v(input_x).view(b, s, n, d)
del input_x
y = self.self_attn.forward(q, k, v, seq_lens)
del q, k, v
if use_token_replace:
x = x + torch.cat([
y[:, :tr_start] * gate_msa,
y[:, tr_start:tr_end] * tr_gate_msa,
y[:, tr_end:] * gate_msa
], dim=1).to(input_dtype)
else:
x = x.addcmul(y, gate_msa)
del y, gate_msa
# ffn
if use_token_replace:
norm2_x = self.norm2(x.to(shift_mlp.dtype))
mod_x = torch.cat([
torch.addcmul(shift_mlp, norm2_x[:, :tr_start], 1 + scale_mlp),
torch.addcmul(tr_shift_mlp, norm2_x[:, tr_start:tr_end], 1 + tr_scale_mlp),
torch.addcmul(shift_mlp, norm2_x[:, tr_end:], 1 + scale_mlp)
], dim=1)
else:
mod_x = torch.addcmul(shift_mlp, self.norm2(x.to(shift_mlp.dtype)), 1 + scale_mlp)
del shift_mlp, scale_mlp
x_ffn = self.ffn_chunked(mod_x.to(input_dtype), num_chunks=1)
del mod_x
# gate_mlp
if use_token_replace:
x = x + torch.cat([
x_ffn[:, :tr_start] * gate_mlp,
x_ffn[:, tr_start:tr_end] * tr_gate_mlp,
x_ffn[:, tr_end:] * gate_mlp
], dim=1).to(input_dtype)
else:
x = x.addcmul(x_ffn.to(gate_mlp.dtype), gate_mlp).to(input_dtype)
del gate_mlp
return x
class WanRefextractor(nn.Module):
def __init__(self, patch_size=(1, 2, 2), in_dim=16, dim=5120, in_features=5120, out_features=5120, ffn_dim=8192, ffn2_dim=8192,
freq_dim=256, num_heads=16, num_layers=32, eps=1e-6,
qk_norm=True, cross_attn_norm=True,
attention_mode='sdpa', rope_func='comfy', rms_norm_function='default',
main_device=torch.device('cuda'), offload_device=torch.device('cpu'), dtype=torch.float16):
super().__init__()
self.patch_size = patch_size
self.freq_dim = freq_dim
self.dim = dim
self.main_device = main_device
self.base_dtype = dtype
self.attention_mode = attention_mode
self.patch_embedding = nn.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size)
self.time_embedding = nn.Sequential(nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim))
self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6))
self.blocks = nn.ModuleList([
WanAttentionBlock(in_features, out_features, ffn_dim, ffn2_dim, num_heads,
qk_norm, cross_attn_norm, eps, attention_mode="sdpa", rope_func=rope_func, rms_norm_function=rms_norm_function)
for i in range(num_layers)
])
self.ref_blocks = nn.ModuleList([])
for _ in range(len(self.blocks)+1):
self.ref_blocks.append(nn.Linear(in_features, out_features))
d = dim // num_heads
self.rope_embedder = EmbedND_RifleX(d,10000.0, [d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)], num_frames=1, k=0)
def rope_encode_comfy(self, t, h, w, freq_offset=0, t_start=0, steps_t=None, steps_h=None, steps_w=None, ntk_alphas=[1,1,1], device=None, dtype=None):
patch_size = self.patch_size
t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
if steps_t is None:
steps_t = t_len
if steps_h is None:
steps_h = h_len
if steps_w is None:
steps_w = w_len
img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start+freq_offset, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(freq_offset, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(freq_offset, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
freqs = self.rope_embedder(img_ids, ntk_alphas).movedim(1, 2)
return freqs
def forward(
self,
x: torch.Tensor,
timestep: torch.LongTensor,
) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
B, C, F, H, W = x.shape
freqs = self.rope_encode_comfy(F, H, W, device=x.device, dtype=x.dtype)
self.patch_embedding.to(self.main_device)
x = self.patch_embedding(x.float()).to(x.dtype).flatten(2).transpose(1, 2).to(self.base_dtype)
seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.int32)
time_embed_dtype = self.time_embedding[0].weight.dtype
if time_embed_dtype not in [torch.float16, torch.bfloat16, torch.float32]:
time_embed_dtype = self.base_dtype
e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, timestep.flatten()).to(time_embed_dtype)) # b, dim
e0 = self.time_projection(e).unflatten(1, (6, self.dim)).to(self.base_dtype) # b, 6, dim
del e
# 4. Transformer blocks
block_samples = ()
for block in self.blocks:
block_samples = block_samples + (x, )
x = block(x, e0, seq_lens, freqs)
block_samples = block_samples + (x, )
ref_block_samples = ()
for block_sample, ref_block in zip(block_samples, self.ref_blocks):
block_sample = ref_block(block_sample)
ref_block_samples = ref_block_samples + (block_sample, )
return ref_block_samples, freqs
|