wehub-resource-sync
1358 lines
48 KiB
Python
1358 lines
48 KiB
Python
# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
|
|
import math
|
|
from typing import Optional
|
|
|
|
import torch
|
|
import torch.cuda.amp as amp
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.models.modeling_utils import ModelMixin
|
|
|
|
from diffusion.model.nets.basic_modules import GLUMBConvTemp, Mlp
|
|
from diffusion.utils.logger import get_logger
|
|
|
|
from .attention import flash_attention
|
|
|
|
__all__ = ["WanModel"]
|
|
|
|
|
|
class AttentionHook:
|
|
output = None
|
|
|
|
def __init__(self, device):
|
|
self.device = device
|
|
|
|
def __call__(self, attn_output):
|
|
self.output = attn_output
|
|
self.output.to(self.device)
|
|
|
|
def clear(self):
|
|
self.output = None
|
|
|
|
|
|
def cosine_similarity(x, y, dim=1):
|
|
x_norm = F.normalize(x, p=2, dim=dim)
|
|
y_norm = F.normalize(y, p=2, dim=dim)
|
|
return torch.sum(x_norm * y_norm, dim=dim)
|
|
|
|
|
|
class BlockHook:
|
|
x_in = None
|
|
x_self_attn = None
|
|
x_cross_attn = None
|
|
x_ffn = None
|
|
|
|
def __init__(self, device, detach=True, score_only="cos"):
|
|
self.device = device
|
|
self.detach = detach if not score_only else True # if score_only, always detach
|
|
self.score_only = score_only
|
|
|
|
def __call__(self, x_in, x_self_attn, x_cross_attn, x_ffn):
|
|
# input shape is B,L,C
|
|
if self.score_only == "cos":
|
|
# make sure all feats are not none
|
|
assert x_in is not None
|
|
assert x_self_attn is not None
|
|
assert x_cross_attn is not None
|
|
assert x_ffn is not None
|
|
# detach and float all feats
|
|
x_in = x_in.detach().float()
|
|
x_self_attn = x_self_attn.detach().float()
|
|
x_cross_attn = x_cross_attn.detach().float()
|
|
x_ffn = x_ffn.detach().float()
|
|
|
|
# compute cosine similarity
|
|
self.x_in = None
|
|
self.x_self_attn = cosine_similarity(x_in, x_self_attn, dim=-1).to(self.device)
|
|
self.x_cross_attn = cosine_similarity(x_self_attn, x_cross_attn, dim=-1).to(self.device)
|
|
self.x_ffn = cosine_similarity(x_cross_attn, x_ffn, dim=-1).to(self.device)
|
|
elif self.score_only == "l2":
|
|
# make sure all feats are not none
|
|
assert x_in is not None
|
|
assert x_self_attn is not None
|
|
assert x_cross_attn is not None
|
|
assert x_ffn is not None
|
|
# detach and float all feats
|
|
x_in = x_in.detach().float()
|
|
x_self_attn = x_self_attn.detach().float()
|
|
x_cross_attn = x_cross_attn.detach().float()
|
|
x_ffn = x_ffn.detach().float()
|
|
|
|
self.x_in = None
|
|
self.x_self_attn = F.mse_loss(x_in, x_self_attn, reduction="none").mean(dim=-1).to(self.device)
|
|
self.x_cross_attn = F.mse_loss(x_self_attn, x_cross_attn, reduction="none").mean(dim=-1).to(self.device)
|
|
self.x_ffn = F.mse_loss(x_cross_attn, x_ffn, reduction="none").mean(dim=-1).to(self.device) # B,L
|
|
else:
|
|
self.x_in = x_in.to(self.device) if x_in is not None else None
|
|
self.x_self_attn = x_self_attn.to(self.device) if x_self_attn is not None else None
|
|
self.x_cross_attn = x_cross_attn.to(self.device) if x_cross_attn is not None else None
|
|
self.x_ffn = x_ffn.to(self.device) if x_ffn is not None else None
|
|
if self.detach:
|
|
self.x_in = self.x_in.detach() if self.x_in is not None else None
|
|
self.x_self_attn = self.x_self_attn.detach() if self.x_self_attn is not None else None
|
|
self.x_cross_attn = self.x_cross_attn.detach() if self.x_cross_attn is not None else None
|
|
self.x_ffn = self.x_ffn.detach() if self.x_ffn is not None else None
|
|
|
|
def clear(self):
|
|
self.x_in = None
|
|
self.x_self_attn = None
|
|
self.x_cross_attn = None
|
|
self.x_ffn = None
|
|
|
|
def get_output(self):
|
|
return {
|
|
"x_in": self.x_in,
|
|
"x_self_attn": self.x_self_attn,
|
|
"x_cross_attn": self.x_cross_attn,
|
|
"x_ffn": self.x_ffn,
|
|
}
|
|
|
|
|
|
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
|
|
|
|
|
|
@amp.autocast(enabled=False)
|
|
def rope_params(max_seq_len, dim, theta=10000):
|
|
assert dim % 2 == 0
|
|
freqs = torch.outer(
|
|
torch.arange(max_seq_len), 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float64).div(dim))
|
|
)
|
|
freqs = torch.polar(torch.ones_like(freqs), freqs)
|
|
return freqs
|
|
|
|
|
|
@amp.autocast(enabled=False)
|
|
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).type_as(x)
|
|
|
|
|
|
class WanSelfAttention(nn.Module):
|
|
def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6, **kwargs):
|
|
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
|
|
|
|
# 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()
|
|
self.qkv_store_buffer = None
|
|
|
|
def forward(self, x, seq_lens, grid_sizes, freqs, block_mask=None):
|
|
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
|
|
# print(f"In Attention, x dtype {x.dtype}")
|
|
x_dtype = x.dtype
|
|
|
|
# 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)
|
|
|
|
q = rope_apply(q, grid_sizes, freqs)
|
|
k = rope_apply(k, grid_sizes, freqs)
|
|
if self.qkv_store_buffer is not None:
|
|
self.qkv_store_buffer["q"] = q[1].cpu() # b, n, h, h_d
|
|
self.qkv_store_buffer["k"] = k[1].cpu() # b, n, h, h_d
|
|
self.qkv_store_buffer["v"] = v[1].cpu() # b, n, h, h_d
|
|
|
|
x = flash_attention(
|
|
q=q,
|
|
k=k,
|
|
v=v,
|
|
k_lens=seq_lens,
|
|
window_size=self.window_size,
|
|
)
|
|
|
|
x = x.flatten(2).to(x_dtype)
|
|
x = self.o(x)
|
|
return x
|
|
|
|
|
|
class WanLinearAttention(WanSelfAttention):
|
|
PAD_VAL = 1
|
|
|
|
def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6, **kwargs):
|
|
super().__init__(dim, num_heads, window_size, qk_norm, eps, **kwargs)
|
|
self.kernel_func = nn.ReLU(inplace=False)
|
|
self.fp32_attention = True
|
|
self.qkv_store_buffer = None
|
|
self.rope_after = kwargs.get("rope_after", False)
|
|
self.power = kwargs.get("power", 1.0)
|
|
|
|
@torch.autocast(device_type="cuda", enabled=False)
|
|
def attn_matmul(self, q, k, v: torch.Tensor) -> torch.Tensor:
|
|
v = F.pad(v.float(), (0, 0, 0, 1), mode="constant", value=self.PAD_VAL)
|
|
vk = torch.matmul(v, k)
|
|
out = torch.matmul(vk, q) # b, h, h_d, n
|
|
|
|
norm_out = out[:, :, :-1] / (out[:, :, -1:] + self.eps)
|
|
|
|
return norm_out
|
|
|
|
def forward(self, x, seq_lens, grid_sizes, freqs):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
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
|
|
x_dtype = x.dtype
|
|
|
|
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) # B, seq, num_heads, head_dim
|
|
|
|
# save before rope
|
|
if self.qkv_store_buffer is not None:
|
|
# qkv store buffer shoud be dict
|
|
self.qkv_store_buffer["q"] = q[1].cpu() # b, n, h, h_d
|
|
self.qkv_store_buffer["k"] = k[1].cpu() # b, n, h, h_d
|
|
self.qkv_store_buffer["v"] = v[1].cpu() # b, n, h, h_d
|
|
|
|
power = self.power
|
|
rope_after = self.rope_after
|
|
if rope_after:
|
|
# apply kernel function
|
|
q = self.kernel_func(q) # B, h, h_d, N
|
|
k = self.kernel_func(k)
|
|
|
|
# power qk
|
|
if power != 1.0:
|
|
q_norm = q.norm(dim=-1, keepdim=True)
|
|
k_norm = k.norm(dim=-1, keepdim=True)
|
|
q = q**power
|
|
k = k**power
|
|
q = (q / (q.norm(dim=-1, keepdim=True) + 1e-6)) * q_norm
|
|
k = (k / (k.norm(dim=-1, keepdim=True) + 1e-6)) * k_norm
|
|
|
|
# apply rope after kernel function
|
|
q_rope = rope_apply(q, grid_sizes, freqs)
|
|
k_rope = rope_apply(k, grid_sizes, freqs)
|
|
|
|
with torch.autocast(device_type="cuda", enabled=False):
|
|
q_rope = q_rope.permute(0, 2, 1, 3).contiguous() # B, seq, num_heads, head_dim -> B, h, seq, h_d
|
|
k_rope = k_rope.permute(0, 2, 1, 3).contiguous() # B, seq, num_heads, head_dim -> B, h, seq, h_d
|
|
|
|
q = q.permute(0, 2, 1, 3).contiguous() # B, seq, num_heads, head_dim -> B, h, seq, h_d
|
|
k = k.permute(0, 2, 1, 3).contiguous() # B, seq, num_heads, head_dim -> B, h, seq, h_d
|
|
v = v.permute(0, 2, 1, 3).contiguous() # B, seq, num_heads, head_dim -> B, h, seq, h_d
|
|
|
|
z = 1 / (q @ k.mean(dim=-2, keepdim=True).transpose(-2, -1) + 1e-6)
|
|
|
|
kv = (k_rope.transpose(-2, -1)) @ v.float() / s
|
|
x = q_rope @ kv
|
|
|
|
x = x * z
|
|
|
|
else:
|
|
# apply rope before kernel function
|
|
q = rope_apply(q, grid_sizes, freqs)
|
|
k = rope_apply(k, grid_sizes, freqs)
|
|
|
|
# apply kernel function
|
|
q = self.kernel_func(q) # B, h, h_d, N
|
|
k = self.kernel_func(k)
|
|
|
|
# power qk
|
|
if power != 1.0:
|
|
q_norm = q.norm(dim=-1, keepdim=True)
|
|
k_norm = k.norm(dim=-1, keepdim=True)
|
|
q = q**power
|
|
k = k**power
|
|
q = (q / (q.norm(dim=-1, keepdim=True) + 1e-6)) * q_norm
|
|
k = (k / (k.norm(dim=-1, keepdim=True) + 1e-6)) * k_norm
|
|
|
|
x = self.attn_matmul(q.permute(0, 2, 3, 1), k.permute(0, 2, 1, 3), v.permute(0, 2, 3, 1))
|
|
|
|
x = x.view(b, n * d, s).permute(0, 2, 1).to(x_dtype) # B, C, N -> B, N, C
|
|
x = self.o(x)
|
|
return x
|
|
|
|
|
|
class STConv(nn.Module):
|
|
def __init__(self, in_dim, out_dim, kernel_size=3, stride=1, padding=1):
|
|
super().__init__()
|
|
self.spatial_conv = nn.Conv2d(in_dim, out_dim, kernel_size, stride, padding, groups=in_dim)
|
|
self.temporal_conv = nn.Conv1d(in_dim, out_dim, kernel_size, stride, padding, groups=in_dim)
|
|
|
|
def forward(self, x):
|
|
B, C, T, H, W = x.shape
|
|
x = x.permute(0, 2, 1, 3, 4).reshape(B * T, C, H, W)
|
|
x = self.spatial_conv(x)
|
|
x = x.reshape(B, T, C, H, W).permute(0, 3, 4, 2, 1).reshape(B * H * W, C, T) # B, T, C, H, W -> B*H*W, C, T
|
|
x = self.temporal_conv(x)
|
|
x = x.reshape(B, H, W, C, T).permute(0, 3, 4, 1, 2).reshape(B, C, T, H, W) # B*H*W, C, T -> B, C, T, H, W
|
|
return x
|
|
|
|
|
|
class MLLALinearAttention(WanLinearAttention):
|
|
def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6, **kwargs):
|
|
super().__init__(dim, num_heads, window_size, qk_norm, eps, **kwargs)
|
|
self.kernel_func = nn.ReLU(inplace=False)
|
|
self.fp32_attention = True
|
|
self.qkv_store_buffer = None
|
|
self.st_conv = STConv(dim, dim)
|
|
self.act = nn.SiLU(inplace=False)
|
|
|
|
def forward(self, x, seq_lens, grid_sizes, freqs):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
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]
|
|
"""
|
|
|
|
F, H, W = grid_sizes[0] # grid size should be the same for all the samples
|
|
B, L, C = x.shape
|
|
|
|
x = x.reshape(B, F, H, W, C).permute(0, 4, 1, 2, 3) # B, C, F, H, W
|
|
x = self.act(self.st_conv(x)).permute(0, 2, 3, 4, 1).reshape(B, L, C)
|
|
|
|
b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
|
|
x_dtype = x.dtype
|
|
|
|
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) # B, seq, num_heads, head_dim
|
|
q = rope_apply(q, grid_sizes, freqs)
|
|
k = rope_apply(k, grid_sizes, freqs)
|
|
|
|
# apply kernel function
|
|
q = self.kernel_func(q) # B, h, h_d, N
|
|
k = self.kernel_func(k)
|
|
# attn matmul input: q: b, h, h_d, n
|
|
# k: b, h, n, h_d
|
|
# v: b, h, h_d, n
|
|
# output: b, h, n, h_d
|
|
x = self.attn_matmul(q.permute(0, 2, 3, 1), k.permute(0, 2, 1, 3), v.permute(0, 2, 3, 1))
|
|
x = x.view(b, n * d, s).permute(0, 2, 1).to(x_dtype) # B, C, N -> B, N, C
|
|
|
|
x = self.o(x) # B, L, C
|
|
|
|
return x
|
|
|
|
|
|
class MLLALePEAttention(WanLinearAttention):
|
|
def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6, **kwargs):
|
|
super().__init__(dim, num_heads, window_size, qk_norm, eps, **kwargs)
|
|
self.kernel_func = nn.ELU(inplace=False)
|
|
self.fp32_attention = True
|
|
self.qkv_store_buffer = None
|
|
self.st_conv = STConv(dim, dim)
|
|
self.act = nn.SiLU(inplace=False)
|
|
self.lepe_conv = STConv(dim, dim)
|
|
|
|
def forward(self, x, seq_lens, grid_sizes, freqs):
|
|
r"""
|
|
Args:
|
|
x(Tensor): Shape [B, L, C]
|
|
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]
|
|
"""
|
|
|
|
F, H, W = grid_sizes[0] # grid size should be the same for all the samples
|
|
B, L, C = x.shape
|
|
|
|
x = x.reshape(B, F, H, W, C).permute(0, 4, 1, 2, 3) # B, C, F, H, W
|
|
x = self.act(self.st_conv(x)).permute(0, 2, 3, 4, 1).reshape(B, L, C)
|
|
|
|
b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
|
|
x_dtype = x.dtype
|
|
|
|
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) # B, seq, num_heads, head_dim
|
|
|
|
# apply kernel function before rope
|
|
q = self.kernel_func(q) + 1
|
|
k = self.kernel_func(k) + 1 # elu + 1
|
|
|
|
# apply rope
|
|
q = rope_apply(q, grid_sizes, freqs)
|
|
k = rope_apply(k, grid_sizes, freqs)
|
|
|
|
# attn matmul input: q: b, h, h_d, n
|
|
# k: b, h, n, h_d
|
|
# v: b, h, h_d, n
|
|
# output: b, h, n, h_d
|
|
x = self.attn_matmul(q.permute(0, 2, 3, 1), k.permute(0, 2, 1, 3), v.permute(0, 2, 3, 1))
|
|
x = x.view(b, n * d, s).permute(0, 2, 1).to(x_dtype) # B, C, N -> B, N, C
|
|
# LePE conv for v
|
|
v = v.reshape(B, F, H, W, C).permute(0, 4, 1, 2, 3) # B, C, F, H, W
|
|
lepe_v = self.lepe_conv(v).permute(0, 2, 3, 4, 1).reshape(B, L, C)
|
|
|
|
x = self.o(x + lepe_v) # B, L, C
|
|
|
|
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 = flash_attention(q, k, v, k_lens=context_lens)
|
|
|
|
# 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):
|
|
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()
|
|
|
|
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 = flash_attention(q, k_img, v_img, k_lens=None)
|
|
# compute attention
|
|
x = flash_attention(q, k, v, k_lens=context_lens)
|
|
|
|
# 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,
|
|
}
|
|
|
|
WAN_SELFATTENTION_CLASSES = {
|
|
"flash": WanSelfAttention,
|
|
"linear": WanLinearAttention,
|
|
"mllalinear": MLLALinearAttention,
|
|
"mllalepe": MLLALePEAttention,
|
|
"bsa": WanSelfAttention,
|
|
}
|
|
|
|
|
|
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,
|
|
self_attn_type="flash",
|
|
rope_after=False,
|
|
power=1.0,
|
|
ffn_type="mlp",
|
|
mlp_acts=("silu", "silu", None),
|
|
):
|
|
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.attn_hook: Optional[AttentionHook] = None
|
|
self.block_hook: Optional[BlockHook] = None
|
|
|
|
# layers
|
|
self.norm1 = WanLayerNorm(dim, eps)
|
|
self.self_attn = WAN_SELFATTENTION_CLASSES[self_attn_type](
|
|
dim, num_heads, window_size, qk_norm, eps, rope_after=rope_after, power=power
|
|
)
|
|
self.self_attn_type = self_attn_type
|
|
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)
|
|
self.norm2 = WanLayerNorm(dim, eps)
|
|
self.ffn = nn.Sequential(nn.Linear(dim, ffn_dim), nn.GELU(approximate="tanh"), nn.Linear(ffn_dim, dim))
|
|
if ffn_type == "mlp":
|
|
self.skip_ffn = None
|
|
elif ffn_type == "GLUMBConvTemp":
|
|
self.skip_ffn = GLUMBConvTemp(
|
|
in_features=dim,
|
|
hidden_features=ffn_dim,
|
|
use_bias=(True, True, False),
|
|
norm=(None, None, None),
|
|
act=mlp_acts,
|
|
t_kernel_size=3,
|
|
)
|
|
nn.init.zeros_(self.skip_ffn.t_conv.weight)
|
|
nn.init.zeros_(self.skip_ffn.point_conv.conv.weight)
|
|
|
|
# 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,
|
|
block_mask=None,
|
|
):
|
|
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
|
|
with amp.autocast(dtype=torch.float32):
|
|
e = (self.modulation + e).chunk(6, dim=1)
|
|
assert e[0].dtype == torch.float32
|
|
|
|
intermediate_feats = {
|
|
"x_in": x,
|
|
"x_self_attn": None,
|
|
"x_cross_attn": None,
|
|
"x_ffn": None,
|
|
}
|
|
# self-attention
|
|
x_dtype = x.dtype
|
|
x_sa_in = (self.norm1(x).float() * (1 + e[1]) + e[0]).to(x_dtype)
|
|
self_attn_kwargs = dict(
|
|
x=x_sa_in,
|
|
seq_lens=seq_lens,
|
|
grid_sizes=grid_sizes,
|
|
freqs=freqs,
|
|
)
|
|
|
|
y = self.self_attn(**self_attn_kwargs)
|
|
|
|
# Call hook if registered
|
|
if self.attn_hook is not None:
|
|
self.attn_hook(y)
|
|
|
|
with amp.autocast(dtype=torch.float32):
|
|
x = x + y * e[2]
|
|
|
|
intermediate_feats["x_self_attn"] = x
|
|
|
|
# cross-attention
|
|
x = x.to(x_dtype)
|
|
x_ca_in = self.norm3(x)
|
|
x = x + self.cross_attn(x_ca_in, context, context_lens)
|
|
|
|
intermediate_feats["x_cross_attn"] = x
|
|
|
|
# ffn
|
|
ffn_in = (self.norm2(x).float() * (1 + e[4]) + e[3]).to(x_dtype)
|
|
y = self.ffn(ffn_in)
|
|
if self.skip_ffn is not None:
|
|
y_skip = self.skip_ffn(ffn_in, HW=grid_sizes[0]) # grid_sizes[0] is three values for T,H,W
|
|
y = y + y_skip
|
|
|
|
with amp.autocast(dtype=torch.float32):
|
|
x = x + y * e[5]
|
|
|
|
intermediate_feats["x_ffn"] = x
|
|
if self.block_hook is not None:
|
|
self.block_hook(**intermediate_feats)
|
|
|
|
del intermediate_feats
|
|
|
|
return x.to(x_dtype)
|
|
|
|
|
|
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]
|
|
"""
|
|
x_type = x.dtype
|
|
assert e.dtype == torch.float32
|
|
with amp.autocast(dtype=torch.float32):
|
|
e = (self.modulation + e.unsqueeze(1)).chunk(2, dim=1)
|
|
x = self.head(self.norm(x) * (1 + e[1]) + e[0])
|
|
return x.to(x_type)
|
|
|
|
|
|
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"]
|
|
ignore_for_config = ["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,
|
|
image_dim=1280,
|
|
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,
|
|
):
|
|
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.gradient_checkpointing = False
|
|
self.enable_autocast = True
|
|
|
|
# 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)
|
|
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
|
|
d = dim // num_heads
|
|
self.freqs = torch.cat(
|
|
[rope_params(1024, d - 4 * (d // 6)), rope_params(1024, 2 * (d // 6)), rope_params(1024, 2 * (d // 6))],
|
|
dim=1,
|
|
)
|
|
|
|
if model_type == "i2v":
|
|
self.img_emb = MLPProj(image_dim, dim)
|
|
|
|
# initialize weights
|
|
self.init_weights()
|
|
self.lr_scale = None
|
|
|
|
def forward(self, x, timestep, context, seq_len, clip_fea=None, y=None, **kwargs):
|
|
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]
|
|
"""
|
|
x = [_x.to(self.dtype) for _x in x]
|
|
context = [_c.to(self.dtype) for _c in context]
|
|
t = timestep
|
|
# params
|
|
device = self.patch_embedding.weight.device
|
|
if self.freqs.device != device:
|
|
self.freqs = self.freqs.to(device)
|
|
|
|
if y is not None:
|
|
x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]
|
|
|
|
# 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))
|
|
assert e.dtype == torch.float32 and e0.dtype == torch.float32
|
|
|
|
# context
|
|
context_lens = None
|
|
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 clip_fea is not None:
|
|
context_clip = self.img_emb(clip_fea) # bs x 257 x dim
|
|
context = torch.concat([context_clip, context], dim=1)
|
|
|
|
# arguments
|
|
kwargs = dict(
|
|
e=e0,
|
|
seq_lens=seq_lens,
|
|
grid_sizes=grid_sizes,
|
|
freqs=self.freqs,
|
|
context=context.to(self.dtype),
|
|
context_lens=context_lens,
|
|
)
|
|
|
|
for i, block in enumerate(self.blocks):
|
|
if self.gradient_checkpointing:
|
|
x = torch.utils.checkpoint.checkpoint(block, x, **kwargs, use_reentrant=False)
|
|
else:
|
|
x = block(x, **kwargs)
|
|
|
|
# head
|
|
x = self.head(x, e)
|
|
|
|
# unpatchify
|
|
x = self.unpatchify(x, grid_sizes)
|
|
return torch.stack(x, dim=0) # .float()
|
|
|
|
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
|
|
out = []
|
|
for u, v in zip(x, grid_sizes.tolist()):
|
|
u = u[: math.prod(v)].view(*v, *self.patch_size, c)
|
|
u = torch.einsum("fhwpqrc->cfphqwr", u)
|
|
u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
|
|
out.append(u)
|
|
return out
|
|
|
|
def init_weights(self):
|
|
r"""
|
|
Initialize model parameters using Xavier initialization.
|
|
"""
|
|
|
|
# basic init
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Linear):
|
|
nn.init.xavier_uniform_(m.weight)
|
|
if m.bias is not None:
|
|
nn.init.zeros_(m.bias)
|
|
|
|
# init embeddings
|
|
nn.init.xavier_uniform_(self.patch_embedding.weight.flatten(1))
|
|
for m in self.text_embedding.modules():
|
|
if isinstance(m, nn.Linear):
|
|
nn.init.normal_(m.weight, std=0.02)
|
|
for m in self.time_embedding.modules():
|
|
if isinstance(m, nn.Linear):
|
|
nn.init.normal_(m.weight, std=0.02)
|
|
|
|
# init output layer
|
|
nn.init.zeros_(self.head.head.weight)
|
|
|
|
# init zero for v_img in each block
|
|
for block in self.blocks:
|
|
if isinstance(block.cross_attn, WanI2VCrossAttention):
|
|
nn.init.zeros_(block.cross_attn.v_img.weight)
|
|
nn.init.zeros_(block.cross_attn.v_img.bias)
|
|
|
|
def load_model_ckpt(self, pretrained_model_path, init_patch_embedding=False, verbose=True, enable_lora=False):
|
|
if enable_lora:
|
|
return self.load_base_model_peft_ckpt(pretrained_model_path, init_patch_embedding, verbose)
|
|
|
|
logger = get_logger(__name__)
|
|
logger.info(f"======> Loading pretrained model {pretrained_model_path} with missing keys <=======")
|
|
if pretrained_model_path.endswith(".safetensors"):
|
|
import safetensors
|
|
|
|
pretrained_model_state_dict = safetensors.torch.load_file(pretrained_model_path, device="cpu")
|
|
elif pretrained_model_path.endswith(".safetensors.index.json"):
|
|
import json
|
|
import os
|
|
|
|
import safetensors
|
|
|
|
index = json.load(open(pretrained_model_path))["weight_map"]
|
|
safetensors_list = set(index.values())
|
|
logger.info(f"======> Loading safetensors {safetensors_list} <=======")
|
|
pretrained_model_state_dict = {}
|
|
for safetensors_path in safetensors_list:
|
|
pretrained_model_state_dict.update(
|
|
safetensors.torch.load_file(
|
|
os.path.join(os.path.dirname(pretrained_model_path), safetensors_path), device="cpu"
|
|
)
|
|
)
|
|
|
|
else:
|
|
pretrained_model_state_dict = torch.load(pretrained_model_path)
|
|
if "state_dict" in pretrained_model_state_dict:
|
|
pretrained_model_state_dict = pretrained_model_state_dict["state_dict"]
|
|
cur_state_dict = self.state_dict()
|
|
new_state_dict = {}
|
|
## load multiview from temporal layer
|
|
non_matched_keys = []
|
|
for k, cur_v in cur_state_dict.items():
|
|
new_state_dict[k] = cur_v
|
|
if k not in pretrained_model_state_dict:
|
|
non_matched_keys.append(k)
|
|
continue
|
|
elif cur_v.shape != pretrained_model_state_dict[k].shape:
|
|
non_matched_keys.append(k)
|
|
continue
|
|
else:
|
|
new_state_dict[k] = pretrained_model_state_dict[k]
|
|
if "patch_embedding.weight" in non_matched_keys:
|
|
# remove patch embedding bias
|
|
non_matched_keys.append("patch_embedding.bias")
|
|
new_state_dict["patch_embedding.bias"] = cur_state_dict["patch_embedding.bias"]
|
|
if init_patch_embedding:
|
|
logger.info("======> init patch embedding <=======")
|
|
non_matched_keys.append("patch_embedding.weight")
|
|
new_state_dict["patch_embedding.weight"] = cur_state_dict["patch_embedding.weight"]
|
|
non_matched_keys.append("patch_embedding.bias")
|
|
new_state_dict["patch_embedding.bias"] = cur_state_dict["patch_embedding.bias"]
|
|
# init head
|
|
non_matched_keys.append("head.head.weight")
|
|
new_state_dict["head.head.weight"] = cur_state_dict["head.head.weight"]
|
|
non_matched_keys.append("head.head.bias")
|
|
new_state_dict["head.head.bias"] = cur_state_dict["head.head.bias"]
|
|
|
|
if verbose:
|
|
for nmk in non_matched_keys:
|
|
logger.warning(f"Non matched key: {nmk}")
|
|
|
|
self.load_state_dict(new_state_dict)
|
|
|
|
def load_state_dict(self, state_dict, strict=True):
|
|
"""Load model with optimizations"""
|
|
from tqdm import tqdm
|
|
|
|
# Convert and move to device in chunks
|
|
logger = get_logger(__name__)
|
|
|
|
chunk_size = 100 # Process 100 parameters at a time
|
|
logger.info(f"Loading model state dict with chunk size {chunk_size}")
|
|
param_items = list(state_dict.items())
|
|
missing_keys = set(self.state_dict().keys()) - set(state_dict.keys())
|
|
unexpected_keys = set(state_dict.keys()) - set(self.state_dict().keys())
|
|
for i in tqdm(range(0, len(param_items), chunk_size)):
|
|
chunk = param_items[i : i + chunk_size]
|
|
|
|
# Process chunk
|
|
for name, tensor in chunk:
|
|
try:
|
|
self.get_parameter(name)
|
|
exists = True
|
|
except:
|
|
exists = False
|
|
if exists:
|
|
# get device and dtype of the parameter
|
|
param_device = self.get_parameter(name).device
|
|
param_dtype = self.get_parameter(name).dtype
|
|
# Move to device with optimal settings
|
|
tensor = tensor.to(device=param_device, dtype=param_dtype, non_blocking=True)
|
|
# Set parameter data
|
|
self.get_parameter(name).data = tensor
|
|
else:
|
|
|
|
if strict:
|
|
raise ValueError(f"Parameter {name} not found in model")
|
|
|
|
return missing_keys, unexpected_keys
|
|
|
|
def register_attn_hook(self, layers=None, device="cpu"):
|
|
for i, block in enumerate(self.blocks):
|
|
if layers is None or i in layers:
|
|
block.attn_hook = AttentionHook(device)
|
|
|
|
def get_attn_output(self):
|
|
attn_outputs = {}
|
|
for i, block in enumerate(self.blocks):
|
|
if block.attn_hook is not None:
|
|
attn_outputs[i] = block.attn_hook.output
|
|
block.attn_hook.clear()
|
|
return attn_outputs
|
|
|
|
def register_block_hook(self, layers=None, device="cpu", detach=True, score_only=False):
|
|
for i, block in enumerate(self.blocks):
|
|
if layers is None or i in layers:
|
|
block.block_hook = BlockHook(device, detach, score_only)
|
|
|
|
def get_block_output(self):
|
|
block_outputs = {}
|
|
for i, block in enumerate(self.blocks):
|
|
if block.block_hook is not None:
|
|
block_outputs[i] = block.block_hook.get_output()
|
|
block.block_hook.clear()
|
|
return block_outputs
|
|
|
|
|
|
class WanLinearAttentionModel(WanModel):
|
|
ignore_for_config = ["cross_attn_norm", "qk_norm", "text_dim", "window_size"]
|
|
|
|
@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,
|
|
image_dim=1280,
|
|
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,
|
|
linear_attn_idx=None,
|
|
attn_type="flash", # flash, linear, mllalinear
|
|
ffn_type="mlp",
|
|
rope_after=False,
|
|
power=1.0,
|
|
):
|
|
super().__init__(
|
|
model_type=model_type,
|
|
patch_size=patch_size,
|
|
text_len=text_len,
|
|
in_dim=in_dim,
|
|
dim=dim,
|
|
ffn_dim=ffn_dim,
|
|
freq_dim=freq_dim,
|
|
image_dim=image_dim,
|
|
text_dim=text_dim,
|
|
out_dim=out_dim,
|
|
num_heads=num_heads,
|
|
num_layers=num_layers,
|
|
window_size=window_size,
|
|
qk_norm=qk_norm,
|
|
cross_attn_norm=cross_attn_norm,
|
|
eps=eps,
|
|
)
|
|
|
|
# blocks
|
|
cross_attn_type = "t2v_cross_attn" if model_type == "t2v" else "i2v_cross_attn"
|
|
self_attn_types = ["flash"] * num_layers
|
|
ffn_types = ["mlp"] * num_layers
|
|
if linear_attn_idx is not None:
|
|
for la_idx in linear_attn_idx:
|
|
self_attn_types[la_idx] = attn_type
|
|
ffn_types[la_idx] = ffn_type
|
|
|
|
self.self_attn_types = self_attn_types
|
|
self.repo_after = rope_after
|
|
self.power = power
|
|
|
|
self.blocks = nn.ModuleList(
|
|
[
|
|
WanAttentionBlock(
|
|
cross_attn_type,
|
|
dim,
|
|
ffn_dim,
|
|
num_heads,
|
|
window_size,
|
|
qk_norm,
|
|
cross_attn_norm,
|
|
eps,
|
|
self_attn_types[i],
|
|
rope_after,
|
|
power,
|
|
ffn_types[i],
|
|
)
|
|
for i in range(num_layers)
|
|
]
|
|
)
|
|
|
|
def forward(self, x, timestep, context, seq_len, clip_fea=None, y=None, block_mask=None, **kwargs):
|
|
r"""
|
|
Forward pass through the diffusion model
|
|
Same as WanModel, but save qkv for linear attention
|
|
|
|
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]
|
|
"""
|
|
x = [_x.to(self.dtype) for _x in x]
|
|
context = [_c.to(self.dtype) for _c in context]
|
|
t = timestep
|
|
self.inference_timestep = int(t[-1].item())
|
|
if not self.training and self.inference_timestep >= 850:
|
|
# NOTE: hard code now. Keep the first several steps using dense attention
|
|
pass
|
|
|
|
if self.model_type == "i2v":
|
|
assert clip_fea is not None and y is not None
|
|
# params
|
|
device = self.patch_embedding.weight.device
|
|
if self.freqs.device != device:
|
|
self.freqs = self.freqs.to(device)
|
|
|
|
if y is not None:
|
|
x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]
|
|
|
|
# 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))
|
|
assert e.dtype == torch.float32 and e0.dtype == torch.float32
|
|
|
|
# context
|
|
context_lens = None
|
|
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 clip_fea is not None:
|
|
context_clip = self.img_emb(clip_fea) # bs x 257 x dim
|
|
context = torch.concat([context_clip, context], dim=1)
|
|
|
|
for i, block in enumerate(self.blocks):
|
|
|
|
# arguments
|
|
kwargs = dict(
|
|
e=e0,
|
|
seq_lens=seq_lens,
|
|
grid_sizes=grid_sizes,
|
|
freqs=self.freqs,
|
|
context=context.to(self.dtype),
|
|
context_lens=context_lens,
|
|
block_mask=None,
|
|
)
|
|
|
|
if self.gradient_checkpointing:
|
|
x = torch.utils.checkpoint.checkpoint(block, x, **kwargs, use_reentrant=False)
|
|
else:
|
|
x = block(x, **kwargs)
|
|
|
|
# head
|
|
x = self.head(x, e)
|
|
|
|
# unpatchify
|
|
x = self.unpatchify(x, grid_sizes)
|
|
|
|
return torch.stack(x, dim=0) # .float()
|
|
|
|
|
|
def init_model_configs(model_cfg, vae_cfg):
|
|
# 1.3B T2V
|
|
model_name = model_cfg.model
|
|
if "1300M" in model_name or "1.3B" in model_name:
|
|
basic_config = {
|
|
"model_type": "t2v",
|
|
"dim": 1536,
|
|
"eps": 1e-06,
|
|
"ffn_dim": 8960,
|
|
"freq_dim": 256,
|
|
"in_dim": 16,
|
|
"num_heads": 12,
|
|
"num_layers": 30,
|
|
"out_dim": 16,
|
|
"text_len": 512,
|
|
"patch_size": (1, 2, 2),
|
|
}
|
|
|
|
elif "14B" in model_name: # default is T2V
|
|
basic_config = {
|
|
"model_type": "t2v",
|
|
"dim": 5120,
|
|
"eps": 1e-06,
|
|
"ffn_dim": 13824,
|
|
"freq_dim": 256,
|
|
"in_dim": 16,
|
|
"num_heads": 40,
|
|
"num_layers": 40,
|
|
"out_dim": 16,
|
|
"text_len": 512,
|
|
"patch_size": (1, 2, 2),
|
|
}
|
|
else:
|
|
raise ValueError(f"Model {model_name} not found")
|
|
|
|
# update basic config with specific model config
|
|
# now all I2V are use cross attention
|
|
if "i2v" in model_name.lower():
|
|
basic_config["model_type"] = "i2v"
|
|
in_dim = vae_cfg.vae_latent_dim * 2
|
|
if model_cfg.mask is not None:
|
|
in_dim += 4
|
|
else:
|
|
in_dim = vae_cfg.vae_latent_dim
|
|
|
|
basic_config["in_dim"] = in_dim
|
|
basic_config["out_dim"] = vae_cfg.vae_latent_dim
|
|
basic_config["patch_size"] = model_cfg.patch_size
|
|
basic_config["linear_attn_idx"] = model_cfg.linear_attn_idx
|
|
basic_config["attn_type"] = model_cfg.self_attn_type
|
|
basic_config["rope_after"] = model_cfg.rope_after
|
|
basic_config["power"] = model_cfg.power
|
|
basic_config["ffn_type"] = model_cfg.ffn_type
|
|
|
|
return basic_config
|