Upload sd_hijack.py
Browse files- sd_hijack.py +195 -0
sd_hijack.py
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import os
|
3 |
+
import sys
|
4 |
+
import traceback
|
5 |
+
import torch
|
6 |
+
import numpy as np
|
7 |
+
from torch import einsum
|
8 |
+
from torch.nn.functional import silu
|
9 |
+
|
10 |
+
import modules.textual_inversion.textual_inversion
|
11 |
+
from modules import prompt_parser, devices, sd_hijack_optimizations, shared, sd_hijack_checkpoint
|
12 |
+
from modules.hypernetworks import hypernetwork
|
13 |
+
from modules.shared import opts, device, cmd_opts
|
14 |
+
from modules import sd_hijack_clip, sd_hijack_open_clip
|
15 |
+
|
16 |
+
from modules.sd_hijack_optimizations import invokeAI_mps_available
|
17 |
+
|
18 |
+
import ldm.modules.attention
|
19 |
+
import ldm.modules.diffusionmodules.model
|
20 |
+
import ldm.models.diffusion.ddim
|
21 |
+
import ldm.models.diffusion.plms
|
22 |
+
import ldm.modules.encoders.modules
|
23 |
+
|
24 |
+
attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
|
25 |
+
diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
|
26 |
+
diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
|
27 |
+
|
28 |
+
# new memory efficient cross attention blocks do not support hypernets and we already
|
29 |
+
# have memory efficient cross attention anyway, so this disables SD2.0's memory efficient cross attention
|
30 |
+
ldm.modules.attention.MemoryEfficientCrossAttention = ldm.modules.attention.CrossAttention
|
31 |
+
ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
|
32 |
+
|
33 |
+
# silence new console spam from SD2
|
34 |
+
ldm.modules.attention.print = lambda *args: None
|
35 |
+
ldm.modules.diffusionmodules.model.print = lambda *args: None
|
36 |
+
|
37 |
+
def apply_optimizations():
|
38 |
+
undo_optimizations()
|
39 |
+
|
40 |
+
ldm.modules.diffusionmodules.model.nonlinearity = silu
|
41 |
+
|
42 |
+
if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and (6, 0) <= torch.cuda.get_device_capability(shared.device) <= (9, 0)):
|
43 |
+
print("Applying xformers cross attention optimization.")
|
44 |
+
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
|
45 |
+
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
|
46 |
+
elif cmd_opts.opt_split_attention_v1:
|
47 |
+
print("Applying v1 cross attention optimization.")
|
48 |
+
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
|
49 |
+
elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()):
|
50 |
+
if not invokeAI_mps_available and shared.device.type == 'mps':
|
51 |
+
print("The InvokeAI cross attention optimization for MPS requires the psutil package which is not installed.")
|
52 |
+
print("Applying v1 cross attention optimization.")
|
53 |
+
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
|
54 |
+
else:
|
55 |
+
print("Applying cross attention optimization (InvokeAI).")
|
56 |
+
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_invokeAI
|
57 |
+
elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):
|
58 |
+
print("Applying cross attention optimization (Doggettx).")
|
59 |
+
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
|
60 |
+
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
|
61 |
+
|
62 |
+
|
63 |
+
def undo_optimizations():
|
64 |
+
ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
|
65 |
+
ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
|
66 |
+
ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
|
67 |
+
|
68 |
+
|
69 |
+
def fix_checkpoint():
|
70 |
+
ldm.modules.attention.BasicTransformerBlock.forward = sd_hijack_checkpoint.BasicTransformerBlock_forward
|
71 |
+
ldm.modules.diffusionmodules.openaimodel.ResBlock.forward = sd_hijack_checkpoint.ResBlock_forward
|
72 |
+
ldm.modules.diffusionmodules.openaimodel.AttentionBlock.forward = sd_hijack_checkpoint.AttentionBlock_forward
|
73 |
+
|
74 |
+
class StableDiffusionModelHijack:
|
75 |
+
fixes = None
|
76 |
+
comments = []
|
77 |
+
layers = None
|
78 |
+
circular_enabled = False
|
79 |
+
clip = None
|
80 |
+
|
81 |
+
embedding_db = modules.textual_inversion.textual_inversion.EmbeddingDatabase(cmd_opts.embeddings_dir)
|
82 |
+
|
83 |
+
def hijack(self, m):
|
84 |
+
if type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenCLIPEmbedder:
|
85 |
+
model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
|
86 |
+
model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
|
87 |
+
m.cond_stage_model = sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
|
88 |
+
elif type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder:
|
89 |
+
m.cond_stage_model.model.token_embedding = EmbeddingsWithFixes(m.cond_stage_model.model.token_embedding, self)
|
90 |
+
m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
|
91 |
+
|
92 |
+
self.clip = m.cond_stage_model
|
93 |
+
|
94 |
+
apply_optimizations()
|
95 |
+
fix_checkpoint()
|
96 |
+
|
97 |
+
def flatten(el):
|
98 |
+
flattened = [flatten(children) for children in el.children()]
|
99 |
+
res = [el]
|
100 |
+
for c in flattened:
|
101 |
+
res += c
|
102 |
+
return res
|
103 |
+
|
104 |
+
self.layers = flatten(m)
|
105 |
+
|
106 |
+
def undo_hijack(self, m):
|
107 |
+
if type(m.cond_stage_model) == sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords:
|
108 |
+
m.cond_stage_model = m.cond_stage_model.wrapped
|
109 |
+
|
110 |
+
model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
|
111 |
+
if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
|
112 |
+
model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
|
113 |
+
elif type(m.cond_stage_model) == sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords:
|
114 |
+
m.cond_stage_model.wrapped.model.token_embedding = m.cond_stage_model.wrapped.model.token_embedding.wrapped
|
115 |
+
m.cond_stage_model = m.cond_stage_model.wrapped
|
116 |
+
|
117 |
+
self.apply_circular(False)
|
118 |
+
self.layers = None
|
119 |
+
self.clip = None
|
120 |
+
|
121 |
+
def apply_circular(self, enable):
|
122 |
+
if self.circular_enabled == enable:
|
123 |
+
return
|
124 |
+
|
125 |
+
self.circular_enabled = enable
|
126 |
+
|
127 |
+
for layer in [layer for layer in self.layers if type(layer) == torch.nn.Conv2d]:
|
128 |
+
layer.padding_mode = 'circular' if enable else 'zeros'
|
129 |
+
|
130 |
+
def clear_comments(self):
|
131 |
+
self.comments = []
|
132 |
+
|
133 |
+
def tokenize(self, text):
|
134 |
+
_, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
|
135 |
+
return remade_batch_tokens[0], token_count, sd_hijack_clip.get_target_prompt_token_count(token_count)
|
136 |
+
|
137 |
+
|
138 |
+
|
139 |
+
class EmbeddingsWithFixes(torch.nn.Module):
|
140 |
+
def __init__(self, wrapped, embeddings):
|
141 |
+
super().__init__()
|
142 |
+
self.wrapped = wrapped
|
143 |
+
self.embeddings = embeddings
|
144 |
+
|
145 |
+
def forward(self, input_ids):
|
146 |
+
batch_fixes = self.embeddings.fixes
|
147 |
+
self.embeddings.fixes = None
|
148 |
+
|
149 |
+
inputs_embeds = self.wrapped(input_ids)
|
150 |
+
|
151 |
+
if batch_fixes is None or len(batch_fixes) == 0 or max([len(x) for x in batch_fixes]) == 0:
|
152 |
+
return inputs_embeds
|
153 |
+
|
154 |
+
vecs = []
|
155 |
+
for fixes, tensor in zip(batch_fixes, inputs_embeds):
|
156 |
+
for offset, embedding in fixes:
|
157 |
+
emb = embedding.vec
|
158 |
+
emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0])
|
159 |
+
tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]])
|
160 |
+
|
161 |
+
vecs.append(tensor)
|
162 |
+
|
163 |
+
return torch.stack(vecs)
|
164 |
+
|
165 |
+
|
166 |
+
def add_circular_option_to_conv_2d():
|
167 |
+
conv2d_constructor = torch.nn.Conv2d.__init__
|
168 |
+
|
169 |
+
def conv2d_constructor_circular(self, *args, **kwargs):
|
170 |
+
return conv2d_constructor(self, *args, padding_mode='circular', **kwargs)
|
171 |
+
|
172 |
+
torch.nn.Conv2d.__init__ = conv2d_constructor_circular
|
173 |
+
|
174 |
+
|
175 |
+
model_hijack = StableDiffusionModelHijack()
|
176 |
+
|
177 |
+
|
178 |
+
def register_buffer(self, name, attr):
|
179 |
+
"""
|
180 |
+
Fix register buffer bug for Mac OS.
|
181 |
+
"""
|
182 |
+
|
183 |
+
if type(attr) == torch.Tensor:
|
184 |
+
if attr.device != devices.device:
|
185 |
+
|
186 |
+
if devices.has_mps():
|
187 |
+
attr = attr.to(device="mps", dtype=torch.float32)
|
188 |
+
else:
|
189 |
+
attr = attr.to(devices.device)
|
190 |
+
|
191 |
+
setattr(self, name, attr)
|
192 |
+
|
193 |
+
|
194 |
+
ldm.models.diffusion.ddim.DDIMSampler.register_buffer = register_buffer
|
195 |
+
ldm.models.diffusion.plms.PLMSSampler.register_buffer = register_buffer
|