458 lines
18 KiB
Python
458 lines
18 KiB
Python
import math
|
|
from dataclasses import dataclass
|
|
|
|
import torch
|
|
from einops import rearrange
|
|
from torch import Tensor, nn
|
|
|
|
from ..math import attention, rope,apply_rope
|
|
|
|
import os
|
|
|
|
class EmbedND(nn.Module):
|
|
def __init__(self, dim: int, theta: int, axes_dim: list[int]):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.theta = theta
|
|
self.axes_dim = axes_dim
|
|
|
|
def forward(self, ids: Tensor) -> Tensor:
|
|
n_axes = ids.shape[-1]
|
|
emb = torch.cat(
|
|
[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
|
|
dim=-3,
|
|
)
|
|
|
|
return emb.unsqueeze(1)
|
|
|
|
|
|
def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
|
|
"""
|
|
Create sinusoidal timestep embeddings.
|
|
:param t: a 1-D Tensor of N indices, one per batch element.
|
|
These may be fractional.
|
|
:param dim: the dimension of the output.
|
|
:param max_period: controls the minimum frequency of the embeddings.
|
|
:return: an (N, D) Tensor of positional embeddings.
|
|
"""
|
|
t = time_factor * t
|
|
half = dim // 2
|
|
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
|
|
t.device
|
|
)
|
|
|
|
args = t[:, None].float() * freqs[None]
|
|
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
|
if dim % 2:
|
|
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
|
if torch.is_floating_point(t):
|
|
embedding = embedding.to(t)
|
|
return embedding
|
|
|
|
|
|
class MLPEmbedder(nn.Module):
|
|
def __init__(self, in_dim: int, hidden_dim: int):
|
|
super().__init__()
|
|
self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
|
|
self.silu = nn.SiLU()
|
|
self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
|
|
|
|
def forward(self, x: Tensor) -> Tensor:
|
|
return self.out_layer(self.silu(self.in_layer(x)))
|
|
|
|
|
|
class RMSNorm(torch.nn.Module):
|
|
def __init__(self, dim: int):
|
|
super().__init__()
|
|
self.scale = nn.Parameter(torch.ones(dim))
|
|
|
|
def forward(self, x: Tensor):
|
|
x_dtype = x.dtype
|
|
x = x.float()
|
|
rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
|
|
return (x * rrms).to(dtype=x_dtype) * self.scale
|
|
|
|
|
|
class QKNorm(torch.nn.Module):
|
|
def __init__(self, dim: int):
|
|
super().__init__()
|
|
self.query_norm = RMSNorm(dim)
|
|
self.key_norm = RMSNorm(dim)
|
|
|
|
def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
|
|
q = self.query_norm(q)
|
|
k = self.key_norm(k)
|
|
return q.to(v), k.to(v)
|
|
|
|
|
|
class SelfAttention(nn.Module):
|
|
def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
|
|
super().__init__()
|
|
self.num_heads = num_heads
|
|
head_dim = dim // num_heads
|
|
|
|
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
|
self.norm = QKNorm(head_dim)
|
|
self.proj = nn.Linear(dim, dim)
|
|
|
|
def forward(self, x: Tensor, pe: Tensor) -> Tensor:
|
|
qkv = self.qkv(x)
|
|
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
q, k = self.norm(q, k, v)
|
|
x = attention(q, k, v, pe=pe)
|
|
x = self.proj(x)
|
|
return x
|
|
|
|
|
|
@dataclass
|
|
class ModulationOut:
|
|
shift: Tensor
|
|
scale: Tensor
|
|
gate: Tensor
|
|
|
|
|
|
class Modulation(nn.Module):
|
|
def __init__(self, dim: int, double: bool):
|
|
super().__init__()
|
|
self.is_double = double
|
|
self.multiplier = 6 if double else 3
|
|
self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
|
|
|
|
def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
|
|
out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
|
|
|
|
return (
|
|
ModulationOut(*out[:3]),
|
|
ModulationOut(*out[3:]) if self.is_double else None,
|
|
)
|
|
|
|
|
|
class DoubleStreamBlock(nn.Module):
|
|
def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
|
|
super().__init__()
|
|
|
|
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
|
self.num_heads = num_heads
|
|
self.hidden_size = hidden_size
|
|
self.img_mod = Modulation(hidden_size, double=True)
|
|
self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
|
|
|
|
self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.img_mlp = nn.Sequential(
|
|
nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
|
|
nn.GELU(approximate="tanh"),
|
|
nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
|
|
)
|
|
|
|
self.txt_mod = Modulation(hidden_size, double=True)
|
|
self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
|
|
|
|
self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.txt_mlp = nn.Sequential(
|
|
nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
|
|
nn.GELU(approximate="tanh"),
|
|
nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
|
|
)
|
|
|
|
def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor) -> tuple[Tensor, Tensor]:
|
|
img_mod1, img_mod2 = self.img_mod(vec)
|
|
txt_mod1, txt_mod2 = self.txt_mod(vec)
|
|
|
|
# prepare image for attention
|
|
img_modulated = self.img_norm1(img)
|
|
img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
|
|
img_qkv = self.img_attn.qkv(img_modulated)
|
|
img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
|
|
img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
|
|
# prepare txt for attention
|
|
txt_modulated = self.txt_norm1(txt)
|
|
txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
|
|
txt_qkv = self.txt_attn.qkv(txt_modulated)
|
|
txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
|
|
|
|
# run actual attention
|
|
q = torch.cat((txt_q, img_q), dim=2)
|
|
k = torch.cat((txt_k, img_k), dim=2)
|
|
v = torch.cat((txt_v, img_v), dim=2)
|
|
# import pdb;pdb.set_trace()
|
|
attn = attention(q, k, v, pe=pe)
|
|
|
|
txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
|
|
|
|
# calculate the img bloks
|
|
img = img + img_mod1.gate * self.img_attn.proj(img_attn)
|
|
img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
|
|
|
|
# calculate the txt bloks
|
|
txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
|
|
txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
|
|
return img, txt
|
|
class SingleStreamBlock(nn.Module):
|
|
"""
|
|
A DiT block with parallel linear layers as described in
|
|
https://arxiv.org/abs/2302.05442 and adapted modulation interface.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
mlp_ratio: float = 4.0,
|
|
qk_scale: float | None = None,
|
|
):
|
|
super().__init__()
|
|
self.hidden_dim = hidden_size
|
|
self.num_heads = num_heads
|
|
head_dim = hidden_size // num_heads
|
|
self.scale = qk_scale or head_dim**-0.5
|
|
|
|
self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
|
# qkv and mlp_in
|
|
self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
|
|
# proj and mlp_out
|
|
self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
|
|
|
|
self.norm = QKNorm(head_dim)
|
|
|
|
self.hidden_size = hidden_size
|
|
self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
|
|
self.mlp_act = nn.GELU(approximate="tanh")
|
|
self.modulation = Modulation(hidden_size, double=False)
|
|
|
|
def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
|
|
mod, _ = self.modulation(vec)
|
|
x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
|
|
qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
|
|
|
|
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
q, k = self.norm(q, k, v)
|
|
|
|
# compute attention
|
|
attn = attention(q, k, v, pe=pe)
|
|
# compute activation in mlp stream, cat again and run second linear layer
|
|
output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
|
|
return x + mod.gate * output
|
|
|
|
|
|
class LastLayer(nn.Module):
|
|
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
|
super().__init__()
|
|
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
|
|
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
|
|
|
|
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
|
shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
|
|
x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
|
|
x = self.linear(x)
|
|
return x
|
|
|
|
class LastLayer(nn.Module):
|
|
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
|
super().__init__()
|
|
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
|
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
|
|
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
|
|
|
|
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
|
shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
|
|
x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
|
|
x = self.linear(x)
|
|
return x
|
|
|
|
import threading
|
|
# 线程局部存储，避免多线程冲突
|
|
_local_storage = threading.local()
|
|
|
|
def get_safe_feature_dict():
|
|
"""获取线程安全的特征字典"""
|
|
if not hasattr(_local_storage, 'safe_feature_dict'):
|
|
_local_storage.safe_feature_dict = {}
|
|
return _local_storage.safe_feature_dict
|
|
|
|
def sync_features_to_safe_dict(info):
|
|
"""将 info['feature'] 中的特征同步到安全字典"""
|
|
if 'feature' in info:
|
|
safe_dict = get_safe_feature_dict()
|
|
for key, value in info['feature'].items():
|
|
if isinstance(value, torch.Tensor):
|
|
# 保存张量的副本，避免设备间干扰
|
|
safe_dict[key] = value.detach() # 保持在当前设备
|
|
else:
|
|
safe_dict[key] = value
|
|
|
|
def get_feature_from_safe_dict(feature_name, target_device):
|
|
"""从安全字典中获取特征"""
|
|
safe_dict = get_safe_feature_dict()
|
|
if feature_name in safe_dict:
|
|
feature = safe_dict[feature_name]
|
|
if isinstance(feature, torch.Tensor):
|
|
return feature.to(target_device)
|
|
return feature
|
|
return None
|
|
|
|
def clear_safe_feature_dict():
|
|
"""清空安全特征字典"""
|
|
if hasattr(_local_storage, 'safe_feature_dict'):
|
|
_local_storage.safe_feature_dict.clear()
|
|
|
|
|
|
class DoubleStreamBlock_kv(DoubleStreamBlock):
|
|
def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
|
|
super().__init__(hidden_size, num_heads, mlp_ratio, qkv_bias)
|
|
|
|
def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, info) -> tuple[Tensor, Tensor]:
|
|
|
|
img_mod1, img_mod2 = self.img_mod(vec)
|
|
txt_mod1, txt_mod2 = self.txt_mod(vec)
|
|
|
|
# prepare image for attention
|
|
img_modulated = self.img_norm1(img)
|
|
img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
|
|
img_qkv = self.img_attn.qkv(img_modulated)
|
|
img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
|
|
img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
|
|
# prepare txt for attention
|
|
txt_modulated = self.txt_norm1(txt)
|
|
txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
|
|
txt_qkv = self.txt_attn.qkv(txt_modulated)
|
|
txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
|
|
|
|
feature_k_name = str(info['t']) + '_' + str(info['id']) + '_' + 'MB' + '_' + 'K'
|
|
feature_v_name = str(info['t']) + '_' + str(info['id']) + '_' + 'MB' + '_' + 'V'
|
|
#print(info['inverse'])
|
|
if info['inverse']:
|
|
info['feature'][feature_k_name] = img_k.cpu()
|
|
info['feature'][feature_v_name] = img_v.cpu()
|
|
|
|
sync_features_to_safe_dict(info)
|
|
|
|
q = torch.cat((txt_q, img_q), dim=2)
|
|
k = torch.cat((txt_k, img_k), dim=2)
|
|
v = torch.cat((txt_v, img_v), dim=2)
|
|
if 'attention_mask' in info:
|
|
if isinstance(info['attention_mask'], torch.Tensor):
|
|
attention_mask_=info['attention_mask'].to(img.device)
|
|
else:
|
|
attention_mask_=info['attention_mask']
|
|
attn = attention(q, k, v, pe=pe,attention_mask=attention_mask_)
|
|
else:
|
|
attn = attention(q, k, v, pe=pe)
|
|
|
|
else:
|
|
if is_feature_available_in_info(info, feature_k_name):
|
|
source_img_k = info['feature'][feature_k_name].to(img.device)
|
|
source_img_v = info['feature'][feature_v_name].to(img.device)
|
|
else:
|
|
source_img_k = get_feature_from_safe_dict(feature_k_name, img.device)
|
|
source_img_v = get_feature_from_safe_dict(feature_v_name, img.device)
|
|
|
|
mask_indices = info['mask_indices']
|
|
source_img_k[:, :, mask_indices, ...] = img_k
|
|
source_img_v[:, :, mask_indices, ...] = img_v
|
|
|
|
q = torch.cat((txt_q, img_q), dim=2)
|
|
k = torch.cat((txt_k, source_img_k), dim=2)
|
|
v = torch.cat((txt_v, source_img_v), dim=2)
|
|
#attn = attention(q, k, v, pe=pe, pe_q = info['pe_mask'])
|
|
attn = attention(q, k, v, pe=pe, pe_q = info['pe_mask'],attention_mask=info['attention_scale'])
|
|
|
|
txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
|
|
|
|
# calculate the img bloks
|
|
img = img + img_mod1.gate * self.img_attn.proj(img_attn)
|
|
img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
|
|
|
|
# calculate the txt bloks
|
|
txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
|
|
txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
|
|
return img, txt
|
|
|
|
def is_feature_available_in_info(info: dict, feature_name: str) -> bool:
|
|
"""检查特征是否在 info 字典中可用"""
|
|
try:
|
|
return (feature_name in info.get('feature', {}) and
|
|
info['feature'][feature_name] is not None)
|
|
except:
|
|
return False
|
|
|
|
class SingleStreamBlock_kv(SingleStreamBlock):
|
|
"""
|
|
A DiT block with parallel linear layers as described in
|
|
https://arxiv.org/abs/2302.05442 and adapted modulation interface.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
num_heads: int,
|
|
mlp_ratio: float = 4.0,
|
|
qk_scale: float | None = None,
|
|
):
|
|
super().__init__(hidden_size, num_heads, mlp_ratio, qk_scale)
|
|
|
|
def forward(self,x: Tensor, vec: Tensor, pe: Tensor, info,) -> Tensor:
|
|
token_len=info['token_len']
|
|
mod, _ = self.modulation(vec)
|
|
x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
|
|
qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
|
|
|
|
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
|
q, k = self.norm(q, k, v)
|
|
# img_k = k[:, :, 512:, ...]
|
|
# img_v = v[:, :, 512:, ...]
|
|
|
|
# txt_k = k[:, :, :512, ...]
|
|
# txt_v = v[:, :, :512, ...]
|
|
img_k = k[:, :, token_len:, ...] #令牌数与clip的必须一致，否则会报错，加载comfyUI的量化T5
|
|
img_v = v[:, :, token_len:, ...]
|
|
|
|
txt_k = k[:, :, :token_len, ...]
|
|
txt_v = v[:, :, :token_len, ...]
|
|
|
|
|
|
feature_k_name = str(info['t']) + '_' + str(info['id']) + '_' + 'SB' + '_' + 'K'
|
|
feature_v_name = str(info['t']) + '_' + str(info['id']) + '_' + 'SB' + '_' + 'V'
|
|
if info['inverse']:
|
|
info['feature'][feature_k_name] = img_k.cpu()
|
|
info['feature'][feature_v_name] = img_v.cpu()
|
|
|
|
sync_features_to_safe_dict(info)
|
|
|
|
if 'attention_mask' in info:
|
|
if isinstance(info['attention_mask'], torch.Tensor):
|
|
attention_mask_=info['attention_mask'].to(x.device)
|
|
else:
|
|
attention_mask_=info['attention_mask']
|
|
attn = attention(q, k, v, pe=pe,attention_mask=attention_mask_)
|
|
else:
|
|
attn = attention(q, k, v, pe=pe)
|
|
|
|
else:
|
|
if is_feature_available_in_info(info, feature_k_name):
|
|
source_img_k = info['feature'][feature_k_name].to(x.device)
|
|
source_img_v = info['feature'][feature_v_name].to(x.device)
|
|
else:
|
|
source_img_k = get_feature_from_safe_dict(feature_k_name, x.device)
|
|
source_img_v = get_feature_from_safe_dict(feature_v_name, x.device)
|
|
|
|
mask_indices = info['mask_indices']
|
|
#print(source_img_k.shape,img_k.shape,mask_indices.shape) #torch.Size([1, 24, 768, 128]) torch.Size([1, 24, 0, 128]) torch.Size([152])
|
|
source_img_k[:, :, mask_indices, ...] = img_k # [1, 24, 154, 128] [24, 0, 128]
|
|
source_img_v[:, :, mask_indices, ...] = img_v
|
|
|
|
k = torch.cat((txt_k, source_img_k), dim=2)
|
|
v = torch.cat((txt_v, source_img_v), dim=2)
|
|
#attn = attention(q, k, v, pe=pe, pe_q = info['pe_mask'])
|
|
attn = attention(q, k, v, pe=pe, pe_q = info['pe_mask'],attention_mask=info['attention_scale'])
|
|
|
|
# compute activation in mlp stream, cat again and run second linear layer
|
|
output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
|
|
return x + mod.gate * output |