Spaces:
Paused
Paused
Upload 2 files
Browse files- uno/flux/modules/conditioner.py +53 -0
- uno/flux/modules/layers.py +435 -0
uno/flux/modules/conditioner.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates. All rights reserved.
|
2 |
+
# Copyright (c) 2024 Black Forest Labs and The XLabs-AI Team. All rights reserved.
|
3 |
+
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
|
16 |
+
from torch import Tensor, nn
|
17 |
+
from transformers import (CLIPTextModel, CLIPTokenizer, T5EncoderModel,
|
18 |
+
T5Tokenizer)
|
19 |
+
|
20 |
+
|
21 |
+
class HFEmbedder(nn.Module):
|
22 |
+
def __init__(self, version: str, max_length: int, **hf_kwargs):
|
23 |
+
super().__init__()
|
24 |
+
self.is_clip = "clip" in version.lower()
|
25 |
+
self.max_length = max_length
|
26 |
+
self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
|
27 |
+
|
28 |
+
if self.is_clip:
|
29 |
+
self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(version, max_length=max_length)
|
30 |
+
self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(version, **hf_kwargs)
|
31 |
+
else:
|
32 |
+
self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(version, max_length=max_length)
|
33 |
+
self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(version, **hf_kwargs)
|
34 |
+
|
35 |
+
self.hf_module = self.hf_module.eval().requires_grad_(False)
|
36 |
+
|
37 |
+
def forward(self, text: list[str]) -> Tensor:
|
38 |
+
batch_encoding = self.tokenizer(
|
39 |
+
text,
|
40 |
+
truncation=True,
|
41 |
+
max_length=self.max_length,
|
42 |
+
return_length=False,
|
43 |
+
return_overflowing_tokens=False,
|
44 |
+
padding="max_length",
|
45 |
+
return_tensors="pt",
|
46 |
+
)
|
47 |
+
|
48 |
+
outputs = self.hf_module(
|
49 |
+
input_ids=batch_encoding["input_ids"].to(self.hf_module.device),
|
50 |
+
attention_mask=None,
|
51 |
+
output_hidden_states=False,
|
52 |
+
)
|
53 |
+
return outputs[self.output_key]
|
uno/flux/modules/layers.py
ADDED
@@ -0,0 +1,435 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates. All rights reserved.
|
2 |
+
# Copyright (c) 2024 Black Forest Labs and The XLabs-AI Team. All rights reserved.
|
3 |
+
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
|
16 |
+
import math
|
17 |
+
from dataclasses import dataclass
|
18 |
+
|
19 |
+
import torch
|
20 |
+
from einops import rearrange
|
21 |
+
from torch import Tensor, nn
|
22 |
+
|
23 |
+
from ..math import attention, rope
|
24 |
+
import torch.nn.functional as F
|
25 |
+
|
26 |
+
class EmbedND(nn.Module):
|
27 |
+
def __init__(self, dim: int, theta: int, axes_dim: list[int]):
|
28 |
+
super().__init__()
|
29 |
+
self.dim = dim
|
30 |
+
self.theta = theta
|
31 |
+
self.axes_dim = axes_dim
|
32 |
+
|
33 |
+
def forward(self, ids: Tensor) -> Tensor:
|
34 |
+
n_axes = ids.shape[-1]
|
35 |
+
emb = torch.cat(
|
36 |
+
[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
|
37 |
+
dim=-3,
|
38 |
+
)
|
39 |
+
|
40 |
+
return emb.unsqueeze(1)
|
41 |
+
|
42 |
+
|
43 |
+
def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
|
44 |
+
"""
|
45 |
+
Create sinusoidal timestep embeddings.
|
46 |
+
:param t: a 1-D Tensor of N indices, one per batch element.
|
47 |
+
These may be fractional.
|
48 |
+
:param dim: the dimension of the output.
|
49 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
50 |
+
:return: an (N, D) Tensor of positional embeddings.
|
51 |
+
"""
|
52 |
+
t = time_factor * t
|
53 |
+
half = dim // 2
|
54 |
+
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
|
55 |
+
t.device
|
56 |
+
)
|
57 |
+
|
58 |
+
args = t[:, None].float() * freqs[None]
|
59 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
60 |
+
if dim % 2:
|
61 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
62 |
+
if torch.is_floating_point(t):
|
63 |
+
embedding = embedding.to(t)
|
64 |
+
return embedding
|
65 |
+
|
66 |
+
|
67 |
+
class MLPEmbedder(nn.Module):
|
68 |
+
def __init__(self, in_dim: int, hidden_dim: int):
|
69 |
+
super().__init__()
|
70 |
+
self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
|
71 |
+
self.silu = nn.SiLU()
|
72 |
+
self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
|
73 |
+
|
74 |
+
def forward(self, x: Tensor) -> Tensor:
|
75 |
+
return self.out_layer(self.silu(self.in_layer(x)))
|
76 |
+
|
77 |
+
|
78 |
+
class RMSNorm(torch.nn.Module):
|
79 |
+
def __init__(self, dim: int):
|
80 |
+
super().__init__()
|
81 |
+
self.scale = nn.Parameter(torch.ones(dim))
|
82 |
+
|
83 |
+
def forward(self, x: Tensor):
|
84 |
+
x_dtype = x.dtype
|
85 |
+
x = x.float()
|
86 |
+
rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
|
87 |
+
return ((x * rrms) * self.scale.float()).to(dtype=x_dtype)
|
88 |
+
|
89 |
+
|
90 |
+
class QKNorm(torch.nn.Module):
|
91 |
+
def __init__(self, dim: int):
|
92 |
+
super().__init__()
|
93 |
+
self.query_norm = RMSNorm(dim)
|
94 |
+
self.key_norm = RMSNorm(dim)
|
95 |
+
|
96 |
+
def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
|
97 |
+
q = self.query_norm(q)
|
98 |
+
k = self.key_norm(k)
|
99 |
+
return q.to(v), k.to(v)
|
100 |
+
|
101 |
+
class LoRALinearLayer(nn.Module):
|
102 |
+
def __init__(self, in_features, out_features, rank=4, network_alpha=None, device=None, dtype=None):
|
103 |
+
super().__init__()
|
104 |
+
|
105 |
+
self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
|
106 |
+
self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
|
107 |
+
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
|
108 |
+
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
|
109 |
+
self.network_alpha = network_alpha
|
110 |
+
self.rank = rank
|
111 |
+
|
112 |
+
nn.init.normal_(self.down.weight, std=1 / rank)
|
113 |
+
nn.init.zeros_(self.up.weight)
|
114 |
+
|
115 |
+
def forward(self, hidden_states):
|
116 |
+
orig_dtype = hidden_states.dtype
|
117 |
+
dtype = self.down.weight.dtype
|
118 |
+
|
119 |
+
down_hidden_states = self.down(hidden_states.to(dtype))
|
120 |
+
up_hidden_states = self.up(down_hidden_states)
|
121 |
+
|
122 |
+
if self.network_alpha is not None:
|
123 |
+
up_hidden_states *= self.network_alpha / self.rank
|
124 |
+
|
125 |
+
return up_hidden_states.to(orig_dtype)
|
126 |
+
|
127 |
+
class FLuxSelfAttnProcessor:
|
128 |
+
def __call__(self, attn, x, pe, **attention_kwargs):
|
129 |
+
qkv = attn.qkv(x)
|
130 |
+
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
131 |
+
q, k = attn.norm(q, k, v)
|
132 |
+
x = attention(q, k, v, pe=pe)
|
133 |
+
x = attn.proj(x)
|
134 |
+
return x
|
135 |
+
|
136 |
+
class LoraFluxAttnProcessor(nn.Module):
|
137 |
+
|
138 |
+
def __init__(self, dim: int, rank=4, network_alpha=None, lora_weight=1):
|
139 |
+
super().__init__()
|
140 |
+
self.qkv_lora = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
|
141 |
+
self.proj_lora = LoRALinearLayer(dim, dim, rank, network_alpha)
|
142 |
+
self.lora_weight = lora_weight
|
143 |
+
|
144 |
+
|
145 |
+
def __call__(self, attn, x, pe, **attention_kwargs):
|
146 |
+
qkv = attn.qkv(x) + self.qkv_lora(x) * self.lora_weight
|
147 |
+
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
|
148 |
+
q, k = attn.norm(q, k, v)
|
149 |
+
x = attention(q, k, v, pe=pe)
|
150 |
+
x = attn.proj(x) + self.proj_lora(x) * self.lora_weight
|
151 |
+
return x
|
152 |
+
|
153 |
+
class SelfAttention(nn.Module):
|
154 |
+
def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
|
155 |
+
super().__init__()
|
156 |
+
self.num_heads = num_heads
|
157 |
+
head_dim = dim // num_heads
|
158 |
+
|
159 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
160 |
+
self.norm = QKNorm(head_dim)
|
161 |
+
self.proj = nn.Linear(dim, dim)
|
162 |
+
def forward():
|
163 |
+
pass
|
164 |
+
|
165 |
+
|
166 |
+
@dataclass
|
167 |
+
class ModulationOut:
|
168 |
+
shift: Tensor
|
169 |
+
scale: Tensor
|
170 |
+
gate: Tensor
|
171 |
+
|
172 |
+
|
173 |
+
class Modulation(nn.Module):
|
174 |
+
def __init__(self, dim: int, double: bool):
|
175 |
+
super().__init__()
|
176 |
+
self.is_double = double
|
177 |
+
self.multiplier = 6 if double else 3
|
178 |
+
self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
|
179 |
+
|
180 |
+
def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
|
181 |
+
out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
|
182 |
+
|
183 |
+
return (
|
184 |
+
ModulationOut(*out[:3]),
|
185 |
+
ModulationOut(*out[3:]) if self.is_double else None,
|
186 |
+
)
|
187 |
+
|
188 |
+
class DoubleStreamBlockLoraProcessor(nn.Module):
|
189 |
+
def __init__(self, dim: int, rank=4, network_alpha=None, lora_weight=1):
|
190 |
+
super().__init__()
|
191 |
+
self.qkv_lora1 = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
|
192 |
+
self.proj_lora1 = LoRALinearLayer(dim, dim, rank, network_alpha)
|
193 |
+
self.qkv_lora2 = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
|
194 |
+
self.proj_lora2 = LoRALinearLayer(dim, dim, rank, network_alpha)
|
195 |
+
self.lora_weight = lora_weight
|
196 |
+
|
197 |
+
def forward(self, attn, img, txt, vec, pe, **attention_kwargs):
|
198 |
+
img_mod1, img_mod2 = attn.img_mod(vec)
|
199 |
+
txt_mod1, txt_mod2 = attn.txt_mod(vec)
|
200 |
+
|
201 |
+
# prepare image for attention
|
202 |
+
img_modulated = attn.img_norm1(img)
|
203 |
+
img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
|
204 |
+
img_qkv = attn.img_attn.qkv(img_modulated) + self.qkv_lora1(img_modulated) * self.lora_weight
|
205 |
+
img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
|
206 |
+
img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)
|
207 |
+
|
208 |
+
# prepare txt for attention
|
209 |
+
txt_modulated = attn.txt_norm1(txt)
|
210 |
+
txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
|
211 |
+
txt_qkv = attn.txt_attn.qkv(txt_modulated) + self.qkv_lora2(txt_modulated) * self.lora_weight
|
212 |
+
txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
|
213 |
+
txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)
|
214 |
+
|
215 |
+
# run actual attention
|
216 |
+
q = torch.cat((txt_q, img_q), dim=2)
|
217 |
+
k = torch.cat((txt_k, img_k), dim=2)
|
218 |
+
v = torch.cat((txt_v, img_v), dim=2)
|
219 |
+
|
220 |
+
attn1 = attention(q, k, v, pe=pe)
|
221 |
+
txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]
|
222 |
+
|
223 |
+
# calculate the img bloks
|
224 |
+
img = img + img_mod1.gate * (attn.img_attn.proj(img_attn) + self.proj_lora1(img_attn) * self.lora_weight)
|
225 |
+
img = img + img_mod2.gate * attn.img_mlp((1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift)
|
226 |
+
|
227 |
+
# calculate the txt bloks
|
228 |
+
txt = txt + txt_mod1.gate * (attn.txt_attn.proj(txt_attn) + self.proj_lora2(txt_attn) * self.lora_weight)
|
229 |
+
txt = txt + txt_mod2.gate * attn.txt_mlp((1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift)
|
230 |
+
return img, txt
|
231 |
+
|
232 |
+
class DoubleStreamBlockProcessor:
|
233 |
+
def __call__(self, attn, img, txt, vec, pe, **attention_kwargs):
|
234 |
+
img_mod1, img_mod2 = attn.img_mod(vec)
|
235 |
+
txt_mod1, txt_mod2 = attn.txt_mod(vec)
|
236 |
+
|
237 |
+
# prepare image for attention
|
238 |
+
img_modulated = attn.img_norm1(img)
|
239 |
+
img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
|
240 |
+
img_qkv = attn.img_attn.qkv(img_modulated)
|
241 |
+
img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads, D=attn.head_dim)
|
242 |
+
img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)
|
243 |
+
|
244 |
+
# prepare txt for attention
|
245 |
+
txt_modulated = attn.txt_norm1(txt)
|
246 |
+
txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
|
247 |
+
txt_qkv = attn.txt_attn.qkv(txt_modulated)
|
248 |
+
txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads, D=attn.head_dim)
|
249 |
+
txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)
|
250 |
+
|
251 |
+
# run actual attention
|
252 |
+
q = torch.cat((txt_q, img_q), dim=2)
|
253 |
+
k = torch.cat((txt_k, img_k), dim=2)
|
254 |
+
v = torch.cat((txt_v, img_v), dim=2)
|
255 |
+
|
256 |
+
attn1 = attention(q, k, v, pe=pe)
|
257 |
+
txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]
|
258 |
+
|
259 |
+
# calculate the img bloks
|
260 |
+
img = img + img_mod1.gate * attn.img_attn.proj(img_attn)
|
261 |
+
img = img + img_mod2.gate * attn.img_mlp((1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift)
|
262 |
+
|
263 |
+
# calculate the txt bloks
|
264 |
+
txt = txt + txt_mod1.gate * attn.txt_attn.proj(txt_attn)
|
265 |
+
txt = txt + txt_mod2.gate * attn.txt_mlp((1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift)
|
266 |
+
return img, txt
|
267 |
+
|
268 |
+
class DoubleStreamBlock(nn.Module):
|
269 |
+
def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
|
270 |
+
super().__init__()
|
271 |
+
mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
272 |
+
self.num_heads = num_heads
|
273 |
+
self.hidden_size = hidden_size
|
274 |
+
self.head_dim = hidden_size // num_heads
|
275 |
+
|
276 |
+
self.img_mod = Modulation(hidden_size, double=True)
|
277 |
+
self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
278 |
+
self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
|
279 |
+
|
280 |
+
self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
281 |
+
self.img_mlp = nn.Sequential(
|
282 |
+
nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
|
283 |
+
nn.GELU(approximate="tanh"),
|
284 |
+
nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
|
285 |
+
)
|
286 |
+
|
287 |
+
self.txt_mod = Modulation(hidden_size, double=True)
|
288 |
+
self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
289 |
+
self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
|
290 |
+
|
291 |
+
self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
292 |
+
self.txt_mlp = nn.Sequential(
|
293 |
+
nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
|
294 |
+
nn.GELU(approximate="tanh"),
|
295 |
+
nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
|
296 |
+
)
|
297 |
+
processor = DoubleStreamBlockProcessor()
|
298 |
+
self.set_processor(processor)
|
299 |
+
|
300 |
+
def set_processor(self, processor) -> None:
|
301 |
+
self.processor = processor
|
302 |
+
|
303 |
+
def get_processor(self):
|
304 |
+
return self.processor
|
305 |
+
|
306 |
+
def forward(
|
307 |
+
self,
|
308 |
+
img: Tensor,
|
309 |
+
txt: Tensor,
|
310 |
+
vec: Tensor,
|
311 |
+
pe: Tensor,
|
312 |
+
image_proj: Tensor = None,
|
313 |
+
ip_scale: float =1.0,
|
314 |
+
) -> tuple[Tensor, Tensor]:
|
315 |
+
if image_proj is None:
|
316 |
+
return self.processor(self, img, txt, vec, pe)
|
317 |
+
else:
|
318 |
+
return self.processor(self, img, txt, vec, pe, image_proj, ip_scale)
|
319 |
+
|
320 |
+
|
321 |
+
class SingleStreamBlockLoraProcessor(nn.Module):
|
322 |
+
def __init__(self, dim: int, rank: int = 4, network_alpha = None, lora_weight: float = 1):
|
323 |
+
super().__init__()
|
324 |
+
self.qkv_lora = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
|
325 |
+
self.proj_lora = LoRALinearLayer(15360, dim, rank, network_alpha)
|
326 |
+
self.lora_weight = lora_weight
|
327 |
+
|
328 |
+
def forward(self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
|
329 |
+
|
330 |
+
mod, _ = attn.modulation(vec)
|
331 |
+
x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
|
332 |
+
qkv, mlp = torch.split(attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1)
|
333 |
+
qkv = qkv + self.qkv_lora(x_mod) * self.lora_weight
|
334 |
+
|
335 |
+
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
|
336 |
+
q, k = attn.norm(q, k, v)
|
337 |
+
|
338 |
+
# compute attention
|
339 |
+
attn_1 = attention(q, k, v, pe=pe)
|
340 |
+
|
341 |
+
# compute activation in mlp stream, cat again and run second linear layer
|
342 |
+
output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
|
343 |
+
output = output + self.proj_lora(torch.cat((attn_1, attn.mlp_act(mlp)), 2)) * self.lora_weight
|
344 |
+
output = x + mod.gate * output
|
345 |
+
return output
|
346 |
+
|
347 |
+
|
348 |
+
class SingleStreamBlockProcessor:
|
349 |
+
def __call__(self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor, **attention_kwargs) -> Tensor:
|
350 |
+
|
351 |
+
mod, _ = attn.modulation(vec)
|
352 |
+
x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
|
353 |
+
qkv, mlp = torch.split(attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1)
|
354 |
+
|
355 |
+
q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
|
356 |
+
q, k = attn.norm(q, k, v)
|
357 |
+
|
358 |
+
# compute attention
|
359 |
+
attn_1 = attention(q, k, v, pe=pe)
|
360 |
+
|
361 |
+
# compute activation in mlp stream, cat again and run second linear layer
|
362 |
+
output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
|
363 |
+
output = x + mod.gate * output
|
364 |
+
return output
|
365 |
+
|
366 |
+
class SingleStreamBlock(nn.Module):
|
367 |
+
"""
|
368 |
+
A DiT block with parallel linear layers as described in
|
369 |
+
https://arxiv.org/abs/2302.05442 and adapted modulation interface.
|
370 |
+
"""
|
371 |
+
|
372 |
+
def __init__(
|
373 |
+
self,
|
374 |
+
hidden_size: int,
|
375 |
+
num_heads: int,
|
376 |
+
mlp_ratio: float = 4.0,
|
377 |
+
qk_scale: float | None = None,
|
378 |
+
):
|
379 |
+
super().__init__()
|
380 |
+
self.hidden_dim = hidden_size
|
381 |
+
self.num_heads = num_heads
|
382 |
+
self.head_dim = hidden_size // num_heads
|
383 |
+
self.scale = qk_scale or self.head_dim**-0.5
|
384 |
+
|
385 |
+
self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
|
386 |
+
# qkv and mlp_in
|
387 |
+
self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
|
388 |
+
# proj and mlp_out
|
389 |
+
self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
|
390 |
+
|
391 |
+
self.norm = QKNorm(self.head_dim)
|
392 |
+
|
393 |
+
self.hidden_size = hidden_size
|
394 |
+
self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
395 |
+
|
396 |
+
self.mlp_act = nn.GELU(approximate="tanh")
|
397 |
+
self.modulation = Modulation(hidden_size, double=False)
|
398 |
+
|
399 |
+
processor = SingleStreamBlockProcessor()
|
400 |
+
self.set_processor(processor)
|
401 |
+
|
402 |
+
|
403 |
+
def set_processor(self, processor) -> None:
|
404 |
+
self.processor = processor
|
405 |
+
|
406 |
+
def get_processor(self):
|
407 |
+
return self.processor
|
408 |
+
|
409 |
+
def forward(
|
410 |
+
self,
|
411 |
+
x: Tensor,
|
412 |
+
vec: Tensor,
|
413 |
+
pe: Tensor,
|
414 |
+
image_proj: Tensor | None = None,
|
415 |
+
ip_scale: float = 1.0,
|
416 |
+
) -> Tensor:
|
417 |
+
if image_proj is None:
|
418 |
+
return self.processor(self, x, vec, pe)
|
419 |
+
else:
|
420 |
+
return self.processor(self, x, vec, pe, image_proj, ip_scale)
|
421 |
+
|
422 |
+
|
423 |
+
|
424 |
+
class LastLayer(nn.Module):
|
425 |
+
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
426 |
+
super().__init__()
|
427 |
+
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
428 |
+
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
|
429 |
+
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
|
430 |
+
|
431 |
+
def forward(self, x: Tensor, vec: Tensor) -> Tensor:
|
432 |
+
shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
|
433 |
+
x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
|
434 |
+
x = self.linear(x)
|
435 |
+
return x
|