833 lines
30 KiB
Python
833 lines
30 KiB
Python
# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
|
|
import math
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.models.modeling_utils import ModelMixin
|
|
|
|
from ...enhance_a_video.enhance import get_feta_scores
|
|
from ...enhance_a_video.globals import is_enhance_enabled
|
|
|
|
from .attention import attention
|
|
import numpy as np
|
|
__all__ = ['WanModel']
|
|
|
|
from tqdm import tqdm
|
|
import gc
|
|
import comfy.model_management as mm
|
|
from ...utils import log, get_module_memory_mb
|
|
|
|
def poly1d(coefficients, x):
|
|
result = torch.zeros_like(x)
|
|
for i, coeff in enumerate(coefficients):
|
|
result += coeff * (x ** (len(coefficients) - 1 - i))
|
|
return result.abs()
|
|
|
|
def sinusoidal_embedding_1d(dim, position):
|
|
# preprocess
|
|
assert dim % 2 == 0
|
|
half = dim // 2
|
|
position = position.type(torch.float64)
|
|
|
|
# calculation
|
|
sinusoid = torch.outer(
|
|
position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
|
|
x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
|
|
return x
|
|
|
|
|
|
def rope_params(max_seq_len, dim, theta=10000, L_test=25, k=0):
|
|
assert dim % 2 == 0
|
|
exponents = torch.arange(0, dim, 2, dtype=torch.float64).div(dim)
|
|
inv_theta_pow = 1.0 / torch.pow(theta, exponents)
|
|
|
|
if k > 0:
|
|
print(f"RifleX: Using {k}th freq")
|
|
inv_theta_pow[k-1] = 0.9 * 2 * torch.pi / L_test
|
|
|
|
freqs = torch.outer(torch.arange(max_seq_len), inv_theta_pow)
|
|
freqs = torch.polar(torch.ones_like(freqs), freqs)
|
|
return freqs
|
|
|
|
from comfy.model_management import get_torch_device, get_autocast_device
|
|
@torch.autocast(device_type=get_autocast_device(get_torch_device()), enabled=False)
|
|
@torch.compiler.disable()
|
|
def rope_apply(x, grid_sizes, freqs):
|
|
n, c = x.size(2), x.size(3) // 2
|
|
|
|
# split freqs
|
|
freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)
|
|
|
|
# loop over samples
|
|
output = []
|
|
for i, (f, h, w) in enumerate(grid_sizes.tolist()):
|
|
seq_len = f * h * w
|
|
|
|
# precompute multipliers
|
|
x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape(
|
|
seq_len, n, -1, 2))
|
|
freqs_i = torch.cat([
|
|
freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
|
|
freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
|
|
freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
|
|
],
|
|
dim=-1).reshape(seq_len, 1, -1)
|
|
|
|
# apply rotary embedding
|
|
x_i = torch.view_as_real(x_i * freqs_i).flatten(2)
|
|
x_i = torch.cat([x_i, x[i, seq_len:]])
|
|
|
|
# append to collection
|
|
output.append(x_i)
|
|
return torch.stack(output).float()
|
|
|
|
|
|
class WanRMSNorm(nn.Module):
|
|
|
|
def __init__(self, dim, eps=1e-5):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.eps = eps
|
|
self.weight = nn.Parameter(torch.ones(dim))
|
|
|
|
def forward(self, x):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
"""
|
|
return self._norm(x.float()).type_as(x) * self.weight
|
|
|
|
def _norm(self, x):
|
|
return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
|
|
|
|
|
|
class WanLayerNorm(nn.LayerNorm):
|
|
|
|
def __init__(self, dim, eps=1e-6, elementwise_affine=False):
|
|
super().__init__(dim, elementwise_affine=elementwise_affine, eps=eps)
|
|
|
|
def forward(self, x):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
"""
|
|
return super().forward(x.float()).type_as(x)
|
|
|
|
|
|
class WanSelfAttention(nn.Module):
|
|
|
|
def __init__(self,
|
|
dim,
|
|
num_heads,
|
|
window_size=(-1, -1),
|
|
qk_norm=True,
|
|
eps=1e-6,
|
|
attention_mode='sdpa'):
|
|
assert dim % num_heads == 0
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.num_heads = num_heads
|
|
self.head_dim = dim // num_heads
|
|
self.window_size = window_size
|
|
self.qk_norm = qk_norm
|
|
self.eps = eps
|
|
self.attention_mode = attention_mode
|
|
|
|
# layers
|
|
self.q = nn.Linear(dim, dim)
|
|
self.k = nn.Linear(dim, dim)
|
|
self.v = nn.Linear(dim, dim)
|
|
self.o = nn.Linear(dim, dim)
|
|
self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
|
|
self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
|
|
|
|
def forward(self, x, seq_lens, grid_sizes, freqs):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, num_heads, C / num_heads]
|
|
seq_lens(Tensor): Shape [B]
|
|
grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
|
|
freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
|
|
"""
|
|
b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
|
|
|
|
# query, key, value function
|
|
def qkv_fn(x):
|
|
q = self.norm_q(self.q(x)).view(b, s, n, d)
|
|
k = self.norm_k(self.k(x)).view(b, s, n, d)
|
|
v = self.v(x).view(b, s, n, d)
|
|
return q, k, v
|
|
|
|
q, k, v = qkv_fn(x)
|
|
|
|
if self.attention_mode == 'spargeattn_tune' or self.attention_mode == 'spargeattn':
|
|
tune_mode = False
|
|
if self.attention_mode == 'spargeattn_tune':
|
|
tune_mode = True
|
|
|
|
if hasattr(self, 'inner_attention'):
|
|
#print("has inner attention")
|
|
q=rope_apply(q, grid_sizes, freqs)
|
|
k=rope_apply(k, grid_sizes, freqs)
|
|
q = q.permute(0, 2, 1, 3)
|
|
k = k.permute(0, 2, 1, 3)
|
|
v = v.permute(0, 2, 1, 3)
|
|
x = self.inner_attention(
|
|
q=q,
|
|
k=k,
|
|
v=v,
|
|
is_causal=False,
|
|
tune_mode=tune_mode
|
|
).permute(0, 2, 1, 3)
|
|
#print("inner attention", x.shape) #inner attention torch.Size([1, 12, 32760, 128])
|
|
else:
|
|
q=rope_apply(q, grid_sizes, freqs)
|
|
k=rope_apply(k, grid_sizes, freqs)
|
|
if is_enhance_enabled():
|
|
feta_scores = get_feta_scores(q, k)
|
|
|
|
x = attention(
|
|
q=q,
|
|
k=k,
|
|
v=v,
|
|
k_lens=seq_lens,
|
|
window_size=self.window_size,
|
|
attention_mode=self.attention_mode)
|
|
|
|
# output
|
|
x = x.flatten(2)
|
|
x = self.o(x)
|
|
|
|
if is_enhance_enabled():
|
|
x *= feta_scores
|
|
|
|
return x
|
|
|
|
|
|
class WanT2VCrossAttention(WanSelfAttention):
|
|
|
|
def forward(self, x, context, context_lens):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L1, C]
|
|
context(Tensor): Shape [B, L2, C]
|
|
context_lens(Tensor): Shape [B]
|
|
"""
|
|
b, n, d = x.size(0), self.num_heads, self.head_dim
|
|
|
|
# compute query, key, value
|
|
q = self.norm_q(self.q(x)).view(b, -1, n, d)
|
|
k = self.norm_k(self.k(context)).view(b, -1, n, d)
|
|
v = self.v(context).view(b, -1, n, d)
|
|
|
|
# compute attention
|
|
x = attention(q, k, v, k_lens=context_lens, attention_mode=self.attention_mode)
|
|
|
|
# output
|
|
x = x.flatten(2)
|
|
x = self.o(x)
|
|
return x
|
|
|
|
|
|
class WanI2VCrossAttention(WanSelfAttention):
|
|
|
|
def __init__(self,
|
|
dim,
|
|
num_heads,
|
|
window_size=(-1, -1),
|
|
qk_norm=True,
|
|
eps=1e-6,
|
|
attention_mode='sdpa'):
|
|
super().__init__(dim, num_heads, window_size, qk_norm, eps)
|
|
|
|
self.k_img = nn.Linear(dim, dim)
|
|
self.v_img = nn.Linear(dim, dim)
|
|
# self.alpha = nn.Parameter(torch.zeros((1, )))
|
|
self.norm_k_img = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
|
|
self.attention_mode = attention_mode
|
|
|
|
def forward(self, x, context, context_lens):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L1, C]
|
|
context(Tensor): Shape [B, L2, C]
|
|
context_lens(Tensor): Shape [B]
|
|
"""
|
|
context_img = context[:, :257]
|
|
context = context[:, 257:]
|
|
b, n, d = x.size(0), self.num_heads, self.head_dim
|
|
|
|
# compute query, key, value
|
|
q = self.norm_q(self.q(x)).view(b, -1, n, d)
|
|
k = self.norm_k(self.k(context)).view(b, -1, n, d)
|
|
v = self.v(context).view(b, -1, n, d)
|
|
k_img = self.norm_k_img(self.k_img(context_img)).view(b, -1, n, d)
|
|
v_img = self.v_img(context_img).view(b, -1, n, d)
|
|
img_x = attention(q, k_img, v_img, k_lens=None, attention_mode=self.attention_mode)
|
|
# compute attention
|
|
x = attention(q, k, v, k_lens=context_lens, attention_mode=self.attention_mode)
|
|
|
|
# output
|
|
x = x.flatten(2)
|
|
img_x = img_x.flatten(2)
|
|
x = x + img_x
|
|
x = self.o(x)
|
|
return x
|
|
|
|
|
|
WAN_CROSSATTENTION_CLASSES = {
|
|
't2v_cross_attn': WanT2VCrossAttention,
|
|
'i2v_cross_attn': WanI2VCrossAttention,
|
|
}
|
|
|
|
|
|
class WanAttentionBlock(nn.Module):
|
|
|
|
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,
|
|
attention_mode='sdpa'):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.ffn_dim = ffn_dim
|
|
self.num_heads = num_heads
|
|
self.window_size = window_size
|
|
self.qk_norm = qk_norm
|
|
self.cross_attn_norm = cross_attn_norm
|
|
self.eps = eps
|
|
self.attention_mode = attention_mode
|
|
|
|
# layers
|
|
self.norm1 = WanLayerNorm(dim, eps)
|
|
self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
|
|
eps, self.attention_mode)
|
|
self.norm3 = WanLayerNorm(
|
|
dim, eps,
|
|
elementwise_affine=True) if cross_attn_norm else nn.Identity()
|
|
self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim,
|
|
num_heads,
|
|
(-1, -1),
|
|
qk_norm,
|
|
eps,#attention_mode=attention_mode sageattn doesn't seem faster here
|
|
)
|
|
self.norm2 = WanLayerNorm(dim, eps)
|
|
self.ffn = nn.Sequential(
|
|
nn.Linear(dim, ffn_dim), nn.GELU(approximate='tanh'),
|
|
nn.Linear(ffn_dim, dim))
|
|
|
|
# modulation
|
|
self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
|
|
|
|
def forward(
|
|
self,
|
|
x,
|
|
e,
|
|
seq_lens,
|
|
grid_sizes,
|
|
freqs,
|
|
context,
|
|
context_lens,
|
|
):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
e(Tensor): Shape [B, 6, C]
|
|
seq_lens(Tensor): Shape [B], length of each sequence in batch
|
|
grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
|
|
freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
|
|
"""
|
|
assert e.dtype == torch.float32
|
|
e = (self.modulation.to(torch.float32).to(e.device) + e.to(torch.float32)).chunk(6, dim=1)
|
|
assert e[0].dtype == torch.float32
|
|
|
|
# self-attention
|
|
y = self.self_attn(
|
|
self.norm1(x).float() * (1 + e[1]) + e[0], seq_lens, grid_sizes,
|
|
freqs)
|
|
x = x.to(torch.float32) + (y.to(torch.float32) * e[2].to(torch.float32))
|
|
|
|
# cross-attention & ffn function
|
|
def cross_attn_ffn(x, context, context_lens, e):
|
|
x = x + self.cross_attn(self.norm3(x), context, context_lens)
|
|
y = self.ffn(self.norm2(x).float() * (1 + e[4]) + e[3])
|
|
x = x.to(torch.float32) + (y.to(torch.float32) * e[5].to(torch.float32))
|
|
return x
|
|
|
|
x = cross_attn_ffn(x, context, context_lens, e)
|
|
return x
|
|
|
|
|
|
class Head(nn.Module):
|
|
|
|
def __init__(self, dim, out_dim, patch_size, eps=1e-6):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.out_dim = out_dim
|
|
self.patch_size = patch_size
|
|
self.eps = eps
|
|
|
|
# layers
|
|
out_dim = math.prod(patch_size) * out_dim
|
|
self.norm = WanLayerNorm(dim, eps)
|
|
self.head = nn.Linear(dim, out_dim)
|
|
|
|
# modulation
|
|
self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5)
|
|
|
|
def forward(self, x, e):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L1, C]
|
|
e(Tensor): Shape [B, C]
|
|
"""
|
|
assert e.dtype == torch.float32
|
|
e_unsqueezed = e.unsqueeze(1).to(torch.float32)
|
|
e = (self.modulation.to(torch.float32).to(e.device) + e_unsqueezed).chunk(2, dim=1)
|
|
normed = self.norm(x).to(torch.float32)
|
|
x = self.head(normed * (1 + e[1].to(torch.float32)) + e[0].to(torch.float32))
|
|
return x
|
|
|
|
|
|
class MLPProj(torch.nn.Module):
|
|
|
|
def __init__(self, in_dim, out_dim):
|
|
super().__init__()
|
|
|
|
self.proj = torch.nn.Sequential(
|
|
torch.nn.LayerNorm(in_dim), torch.nn.Linear(in_dim, in_dim),
|
|
torch.nn.GELU(), torch.nn.Linear(in_dim, out_dim),
|
|
torch.nn.LayerNorm(out_dim))
|
|
|
|
def forward(self, image_embeds):
|
|
clip_extra_context_tokens = self.proj(image_embeds)
|
|
return clip_extra_context_tokens
|
|
|
|
|
|
class WanModel(ModelMixin, ConfigMixin):
|
|
r"""
|
|
Wan diffusion backbone supporting both text-to-video and image-to-video.
|
|
"""
|
|
|
|
ignore_for_config = [
|
|
'patch_size', 'cross_attn_norm', 'qk_norm', 'text_dim', 'window_size'
|
|
]
|
|
_no_split_modules = ['WanAttentionBlock']
|
|
|
|
@register_to_config
|
|
def __init__(self,
|
|
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,
|
|
attention_mode='sdpa',
|
|
main_device=torch.device('cuda'),
|
|
offload_device=torch.device('cpu'),
|
|
teacache_coefficients=[],):
|
|
r"""
|
|
Initialize the diffusion model backbone.
|
|
|
|
Args:
|
|
model_type (`str`, *optional*, defaults to 't2v'):
|
|
Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
|
|
patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
|
|
3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
|
|
text_len (`int`, *optional*, defaults to 512):
|
|
Fixed length for text embeddings
|
|
in_dim (`int`, *optional*, defaults to 16):
|
|
Input video channels (C_in)
|
|
dim (`int`, *optional*, defaults to 2048):
|
|
Hidden dimension of the transformer
|
|
ffn_dim (`int`, *optional*, defaults to 8192):
|
|
Intermediate dimension in feed-forward network
|
|
freq_dim (`int`, *optional*, defaults to 256):
|
|
Dimension for sinusoidal time embeddings
|
|
text_dim (`int`, *optional*, defaults to 4096):
|
|
Input dimension for text embeddings
|
|
out_dim (`int`, *optional*, defaults to 16):
|
|
Output video channels (C_out)
|
|
num_heads (`int`, *optional*, defaults to 16):
|
|
Number of attention heads
|
|
num_layers (`int`, *optional*, defaults to 32):
|
|
Number of transformer blocks
|
|
window_size (`tuple`, *optional*, defaults to (-1, -1)):
|
|
Window size for local attention (-1 indicates global attention)
|
|
qk_norm (`bool`, *optional*, defaults to True):
|
|
Enable query/key normalization
|
|
cross_attn_norm (`bool`, *optional*, defaults to False):
|
|
Enable cross-attention normalization
|
|
eps (`float`, *optional*, defaults to 1e-6):
|
|
Epsilon value for normalization layers
|
|
"""
|
|
|
|
super().__init__()
|
|
|
|
assert model_type in ['t2v', 'i2v']
|
|
self.model_type = model_type
|
|
|
|
self.patch_size = patch_size
|
|
self.text_len = text_len
|
|
self.in_dim = in_dim
|
|
self.dim = dim
|
|
self.ffn_dim = ffn_dim
|
|
self.freq_dim = freq_dim
|
|
self.text_dim = text_dim
|
|
self.out_dim = out_dim
|
|
self.num_heads = num_heads
|
|
self.num_layers = num_layers
|
|
self.window_size = window_size
|
|
self.qk_norm = qk_norm
|
|
self.cross_attn_norm = cross_attn_norm
|
|
self.eps = eps
|
|
self.attention_mode = attention_mode
|
|
self.main_device = main_device
|
|
self.offload_device = offload_device
|
|
|
|
self.blocks_to_swap = -1
|
|
self.offload_txt_emb = False
|
|
self.offload_img_emb = False
|
|
|
|
#init TeaCache variables
|
|
self.enable_teacache = False
|
|
self.rel_l1_thresh = 0.15
|
|
self.teacache_start_step= 0
|
|
self.teacache_end_step = -1
|
|
self.teacache_cache_device = main_device
|
|
self.teacache_state = TeaCacheState()
|
|
self.teacache_coefficients = teacache_coefficients
|
|
self.teacache_use_coefficients = False
|
|
# self.l1_history_x = []
|
|
# self.l1_history_temb = []
|
|
# self.l1_history_rescaled = []
|
|
|
|
# embeddings
|
|
self.patch_embedding = nn.Conv3d(
|
|
in_dim, dim, kernel_size=patch_size, stride=patch_size)
|
|
self.text_embedding = nn.Sequential(
|
|
nn.Linear(text_dim, dim), nn.GELU(approximate='tanh'),
|
|
nn.Linear(dim, dim))
|
|
|
|
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))
|
|
|
|
# blocks
|
|
cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
|
|
self.blocks = nn.ModuleList([
|
|
WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
|
|
window_size, qk_norm, cross_attn_norm, eps,
|
|
attention_mode=self.attention_mode)
|
|
for _ in range(num_layers)
|
|
])
|
|
|
|
# head
|
|
self.head = Head(dim, out_dim, patch_size, eps)
|
|
|
|
# buffers (don't use register_buffer otherwise dtype will be changed in to())
|
|
assert (dim % num_heads) == 0 and (dim // num_heads) % 2 == 0
|
|
|
|
|
|
if model_type == 'i2v':
|
|
self.img_emb = MLPProj(1280, dim)
|
|
|
|
# initialize weights
|
|
#self.init_weights()
|
|
|
|
def block_swap(self, blocks_to_swap, offload_txt_emb=False, offload_img_emb=False):
|
|
print(f"Swapping {blocks_to_swap + 1} transformer blocks")
|
|
self.blocks_to_swap = blocks_to_swap
|
|
self.offload_img_emb = offload_img_emb
|
|
self.offload_txt_emb = offload_txt_emb
|
|
|
|
total_offload_memory = 0
|
|
total_main_memory = 0
|
|
|
|
for b, block in tqdm(enumerate(self.blocks), total=len(self.blocks), desc="Initializing block swap"):
|
|
block_memory = get_module_memory_mb(block)
|
|
|
|
if b > self.blocks_to_swap:
|
|
block.to(self.main_device)
|
|
total_main_memory += block_memory
|
|
else:
|
|
block.to(self.offload_device)
|
|
total_offload_memory += block_memory
|
|
|
|
mm.soft_empty_cache()
|
|
gc.collect()
|
|
|
|
#print(f"Block {b}: {block_memory:.2f}MB on {block.parameters().__next__().device}")
|
|
log.info("----------------------")
|
|
log.info(f"Block swap memory summary:")
|
|
log.info(f"Transformer blocks on {self.offload_device}: {total_offload_memory:.2f}MB")
|
|
log.info(f"Transformer blocks on {self.main_device}: {total_main_memory:.2f}MB")
|
|
log.info(f"Total memory used by transformer blocks: {(total_offload_memory + total_main_memory):.2f}MB")
|
|
log.info("----------------------")
|
|
|
|
def forward(
|
|
self,
|
|
x,
|
|
t,
|
|
context,
|
|
seq_len,
|
|
clip_fea=None,
|
|
y=None,
|
|
device=torch.device('cuda'),
|
|
freqs=None,
|
|
current_step=0,
|
|
pred_id=None
|
|
):
|
|
r"""
|
|
Forward pass through the diffusion model
|
|
|
|
Args:
|
|
x (List[Tensor]):
|
|
List of input video tensors, each with shape [C_in, F, H, W]
|
|
t (Tensor):
|
|
Diffusion timesteps tensor of shape [B]
|
|
context (List[Tensor]):
|
|
List of text embeddings each with shape [L, C]
|
|
seq_len (`int`):
|
|
Maximum sequence length for positional encoding
|
|
clip_fea (Tensor, *optional*):
|
|
CLIP image features for image-to-video mode
|
|
y (List[Tensor], *optional*):
|
|
Conditional video inputs for image-to-video mode, same shape as x
|
|
|
|
Returns:
|
|
List[Tensor]:
|
|
List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
|
|
"""
|
|
if self.model_type == 'i2v':
|
|
assert clip_fea is not None and y is not None
|
|
# params
|
|
#device = self.patch_embedding.weight.device
|
|
if freqs.device != device:
|
|
freqs = freqs.to(device)
|
|
|
|
if y is not None:
|
|
x = torch.cat([x, y], dim=0)
|
|
|
|
# embeddings
|
|
x = [self.patch_embedding(x.unsqueeze(0))]
|
|
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 torch.autocast(device_type='cuda', 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))
|
|
assert e.dtype == torch.float32 and e0.dtype == torch.float32
|
|
|
|
# context
|
|
context_lens = None
|
|
if self.offload_txt_emb:
|
|
self.text_embedding.to(self.main_device)
|
|
context = self.text_embedding(
|
|
torch.stack([
|
|
torch.cat(
|
|
[u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
|
|
for u in context
|
|
]))
|
|
if self.offload_txt_emb:
|
|
self.text_embedding.to(self.offload_device, non_blocking=True)
|
|
|
|
if clip_fea is not None:
|
|
if self.offload_img_emb:
|
|
self.img_emb.to(self.main_device)
|
|
context_clip = self.img_emb(clip_fea) # bs x 257 x dim
|
|
context = torch.concat([context_clip, context], dim=1)
|
|
if self.offload_img_emb:
|
|
self.img_emb.to(self.offload_device, non_blocking=True)
|
|
|
|
should_calc = True
|
|
accumulated_rel_l1_distance = torch.tensor(0.0, dtype=torch.float32, device=device)
|
|
if self.enable_teacache and self.teacache_start_step <= current_step <= self.teacache_end_step:
|
|
if pred_id is None:
|
|
pred_id = self.teacache_state.new_prediction()
|
|
#log.info(current_step)
|
|
#log.info(f"TeaCache: Initializing TeaCache variables for model pred: {pred_id}")
|
|
should_calc = True
|
|
else:
|
|
previous_modulated_input = self.teacache_state.get(pred_id)['previous_modulated_input']
|
|
previous_modulated_input = previous_modulated_input.to(device)
|
|
previous_residual = self.teacache_state.get(pred_id)['previous_residual']
|
|
accumulated_rel_l1_distance = self.teacache_state.get(pred_id)['accumulated_rel_l1_distance']
|
|
|
|
if self.teacache_use_coefficients:
|
|
rescale_func = np.poly1d(self.teacache_coefficients)
|
|
accumulated_rel_l1_distance += rescale_func(((e-previous_modulated_input).abs().mean() / previous_modulated_input.abs().mean()).cpu().item())
|
|
else:
|
|
temb_relative_l1 = relative_l1_distance(previous_modulated_input, e0)
|
|
accumulated_rel_l1_distance = accumulated_rel_l1_distance.to(e0.device) + temb_relative_l1
|
|
|
|
#print("accumulated_rel_l1_distance", accumulated_rel_l1_distance)
|
|
|
|
if accumulated_rel_l1_distance < self.rel_l1_thresh:
|
|
should_calc = False
|
|
else:
|
|
should_calc = True
|
|
accumulated_rel_l1_distance = torch.tensor(0.0, dtype=torch.float32, device=device)
|
|
|
|
previous_modulated_input = e.clone() if self.teacache_use_coefficients else e0.clone()
|
|
if not should_calc:
|
|
x += previous_residual.to(x.device)
|
|
#log.info(f"TeaCache: Skipping uncond step {current_step+1}")
|
|
self.teacache_state.update(
|
|
pred_id,
|
|
accumulated_rel_l1_distance=accumulated_rel_l1_distance,
|
|
skipped_steps=self.teacache_state.get(pred_id)['skipped_steps'] + 1,
|
|
)
|
|
|
|
if not self.enable_teacache or (self.enable_teacache and should_calc):
|
|
if self.enable_teacache:
|
|
original_x = x.clone()
|
|
# arguments
|
|
kwargs = dict(
|
|
e=e0,
|
|
seq_lens=seq_lens,
|
|
grid_sizes=grid_sizes,
|
|
freqs=freqs,
|
|
context=context,
|
|
context_lens=context_lens)
|
|
|
|
for b, block in enumerate(self.blocks):
|
|
if b <= self.blocks_to_swap and self.blocks_to_swap >= 0:
|
|
block.to(self.main_device)
|
|
x = block(x, **kwargs)
|
|
if b <= self.blocks_to_swap and self.blocks_to_swap >= 0:
|
|
block.to(self.offload_device, non_blocking=True)
|
|
|
|
if self.enable_teacache and pred_id is not None:
|
|
self.teacache_state.update(
|
|
pred_id,
|
|
previous_residual=(x - original_x),
|
|
accumulated_rel_l1_distance=accumulated_rel_l1_distance,
|
|
previous_modulated_input=previous_modulated_input
|
|
)
|
|
#self.teacache_state.report()
|
|
|
|
# head
|
|
x = self.head(x, e)
|
|
# unpatchify
|
|
x = self.unpatchify(x, grid_sizes)
|
|
return x, pred_id
|
|
|
|
def unpatchify(self, x, grid_sizes):
|
|
r"""
|
|
Reconstruct video tensors from patch embeddings.
|
|
|
|
Args:
|
|
x (List[Tensor]):
|
|
List of patchified features, each with shape [L, C_out * prod(patch_size)]
|
|
grid_sizes (Tensor):
|
|
Original spatial-temporal grid dimensions before patching,
|
|
shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
|
|
|
|
Returns:
|
|
List[Tensor]:
|
|
Reconstructed video tensors with shape [C_out, F, H / 8, W / 8]
|
|
"""
|
|
|
|
c = self.out_dim
|
|
for v in grid_sizes.tolist():
|
|
x = x[:math.prod(v)].view(*v, *self.patch_size, c)
|
|
x = torch.einsum('fhwpqrc->cfphqwr', x)
|
|
x = x.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
|
|
return x
|
|
|
|
class TeaCacheState:
|
|
def __init__(self, cache_device='cpu'):
|
|
self.cache_device = cache_device
|
|
self.states = {}
|
|
self._next_pred_id = 0
|
|
|
|
def new_prediction(self):
|
|
"""Create new prediction state and return its ID"""
|
|
pred_id = self._next_pred_id
|
|
self._next_pred_id += 1
|
|
self.states[pred_id] = {
|
|
'previous_residual': None,
|
|
'accumulated_rel_l1_distance': 0,
|
|
'previous_modulated_input': None,
|
|
'skipped_steps': 0
|
|
}
|
|
return pred_id
|
|
|
|
def update(self, pred_id, **kwargs):
|
|
"""Update state for specific prediction"""
|
|
if pred_id not in self.states:
|
|
return None
|
|
for key, value in kwargs.items():
|
|
if isinstance(value, torch.Tensor):
|
|
value = value.to(self.cache_device)
|
|
self.states[pred_id][key] = value
|
|
|
|
def get(self, pred_id):
|
|
return self.states.get(pred_id, {})
|
|
|
|
def report(self):
|
|
for pred_id in self.states:
|
|
log.info(f"Prediction {pred_id}: {self.states[pred_id]}")
|
|
|
|
def clear_prediction(self, pred_id):
|
|
if pred_id in self.states:
|
|
del self.states[pred_id]
|
|
|
|
def clear_all(self):
|
|
self.states.clear()
|
|
self._next_pred_id = 0
|
|
|
|
def relative_l1_distance(last_tensor, current_tensor):
|
|
l1_distance = torch.abs(last_tensor.to(current_tensor.device) - current_tensor).mean()
|
|
norm = torch.abs(last_tensor).mean()
|
|
relative_l1_distance = l1_distance / norm
|
|
return relative_l1_distance.to(torch.float32).to(current_tensor.device)
|
|
|
|
def normalize_values(values):
|
|
min_val = min(values)
|
|
max_val = max(values)
|
|
if max_val == min_val:
|
|
return [0.0] * len(values)
|
|
return [(x - min_val) / (max_val - min_val) for x in values]
|
|
|
|
def rescale_differences(input_diffs, output_diffs):
|
|
"""Polynomial fitting between input and output differences"""
|
|
poly_degree = 4
|
|
if len(input_diffs) < 2:
|
|
return input_diffs
|
|
|
|
x = np.array([x.item() for x in input_diffs])
|
|
y = np.array([y.item() for y in output_diffs])
|
|
print("x ", x)
|
|
print("y ", y)
|
|
|
|
# Fit polynomial
|
|
coeffs = np.polyfit(x, y, poly_degree)
|
|
|
|
# Apply polynomial transformation
|
|
return np.polyval(coeffs, x) |