Upload lora-scripts/sd-scripts/library/attention_processors.py with huggingface_hub
Browse files
lora-scripts/sd-scripts/library/attention_processors.py
ADDED
@@ -0,0 +1,227 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import Any
|
3 |
+
from einops import rearrange
|
4 |
+
import torch
|
5 |
+
from diffusers.models.attention_processor import Attention
|
6 |
+
|
7 |
+
|
8 |
+
# flash attention forwards and backwards
|
9 |
+
|
10 |
+
# https://arxiv.org/abs/2205.14135
|
11 |
+
|
12 |
+
EPSILON = 1e-6
|
13 |
+
|
14 |
+
|
15 |
+
class FlashAttentionFunction(torch.autograd.function.Function):
|
16 |
+
@staticmethod
|
17 |
+
@torch.no_grad()
|
18 |
+
def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
|
19 |
+
"""Algorithm 2 in the paper"""
|
20 |
+
|
21 |
+
device = q.device
|
22 |
+
dtype = q.dtype
|
23 |
+
max_neg_value = -torch.finfo(q.dtype).max
|
24 |
+
qk_len_diff = max(k.shape[-2] - q.shape[-2], 0)
|
25 |
+
|
26 |
+
o = torch.zeros_like(q)
|
27 |
+
all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device)
|
28 |
+
all_row_maxes = torch.full(
|
29 |
+
(*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device
|
30 |
+
)
|
31 |
+
|
32 |
+
scale = q.shape[-1] ** -0.5
|
33 |
+
|
34 |
+
if mask is None:
|
35 |
+
mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size)
|
36 |
+
else:
|
37 |
+
mask = rearrange(mask, "b n -> b 1 1 n")
|
38 |
+
mask = mask.split(q_bucket_size, dim=-1)
|
39 |
+
|
40 |
+
row_splits = zip(
|
41 |
+
q.split(q_bucket_size, dim=-2),
|
42 |
+
o.split(q_bucket_size, dim=-2),
|
43 |
+
mask,
|
44 |
+
all_row_sums.split(q_bucket_size, dim=-2),
|
45 |
+
all_row_maxes.split(q_bucket_size, dim=-2),
|
46 |
+
)
|
47 |
+
|
48 |
+
for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits):
|
49 |
+
q_start_index = ind * q_bucket_size - qk_len_diff
|
50 |
+
|
51 |
+
col_splits = zip(
|
52 |
+
k.split(k_bucket_size, dim=-2),
|
53 |
+
v.split(k_bucket_size, dim=-2),
|
54 |
+
)
|
55 |
+
|
56 |
+
for k_ind, (kc, vc) in enumerate(col_splits):
|
57 |
+
k_start_index = k_ind * k_bucket_size
|
58 |
+
|
59 |
+
attn_weights = (
|
60 |
+
torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale
|
61 |
+
)
|
62 |
+
|
63 |
+
if row_mask is not None:
|
64 |
+
attn_weights.masked_fill_(~row_mask, max_neg_value)
|
65 |
+
|
66 |
+
if causal and q_start_index < (k_start_index + k_bucket_size - 1):
|
67 |
+
causal_mask = torch.ones(
|
68 |
+
(qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device
|
69 |
+
).triu(q_start_index - k_start_index + 1)
|
70 |
+
attn_weights.masked_fill_(causal_mask, max_neg_value)
|
71 |
+
|
72 |
+
block_row_maxes = attn_weights.amax(dim=-1, keepdims=True)
|
73 |
+
attn_weights -= block_row_maxes
|
74 |
+
exp_weights = torch.exp(attn_weights)
|
75 |
+
|
76 |
+
if row_mask is not None:
|
77 |
+
exp_weights.masked_fill_(~row_mask, 0.0)
|
78 |
+
|
79 |
+
block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp(
|
80 |
+
min=EPSILON
|
81 |
+
)
|
82 |
+
|
83 |
+
new_row_maxes = torch.maximum(block_row_maxes, row_maxes)
|
84 |
+
|
85 |
+
exp_values = torch.einsum(
|
86 |
+
"... i j, ... j d -> ... i d", exp_weights, vc
|
87 |
+
)
|
88 |
+
|
89 |
+
exp_row_max_diff = torch.exp(row_maxes - new_row_maxes)
|
90 |
+
exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes)
|
91 |
+
|
92 |
+
new_row_sums = (
|
93 |
+
exp_row_max_diff * row_sums
|
94 |
+
+ exp_block_row_max_diff * block_row_sums
|
95 |
+
)
|
96 |
+
|
97 |
+
oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_(
|
98 |
+
(exp_block_row_max_diff / new_row_sums) * exp_values
|
99 |
+
)
|
100 |
+
|
101 |
+
row_maxes.copy_(new_row_maxes)
|
102 |
+
row_sums.copy_(new_row_sums)
|
103 |
+
|
104 |
+
ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size)
|
105 |
+
ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes)
|
106 |
+
|
107 |
+
return o
|
108 |
+
|
109 |
+
@staticmethod
|
110 |
+
@torch.no_grad()
|
111 |
+
def backward(ctx, do):
|
112 |
+
"""Algorithm 4 in the paper"""
|
113 |
+
|
114 |
+
causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args
|
115 |
+
q, k, v, o, l, m = ctx.saved_tensors
|
116 |
+
|
117 |
+
device = q.device
|
118 |
+
|
119 |
+
max_neg_value = -torch.finfo(q.dtype).max
|
120 |
+
qk_len_diff = max(k.shape[-2] - q.shape[-2], 0)
|
121 |
+
|
122 |
+
dq = torch.zeros_like(q)
|
123 |
+
dk = torch.zeros_like(k)
|
124 |
+
dv = torch.zeros_like(v)
|
125 |
+
|
126 |
+
row_splits = zip(
|
127 |
+
q.split(q_bucket_size, dim=-2),
|
128 |
+
o.split(q_bucket_size, dim=-2),
|
129 |
+
do.split(q_bucket_size, dim=-2),
|
130 |
+
mask,
|
131 |
+
l.split(q_bucket_size, dim=-2),
|
132 |
+
m.split(q_bucket_size, dim=-2),
|
133 |
+
dq.split(q_bucket_size, dim=-2),
|
134 |
+
)
|
135 |
+
|
136 |
+
for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits):
|
137 |
+
q_start_index = ind * q_bucket_size - qk_len_diff
|
138 |
+
|
139 |
+
col_splits = zip(
|
140 |
+
k.split(k_bucket_size, dim=-2),
|
141 |
+
v.split(k_bucket_size, dim=-2),
|
142 |
+
dk.split(k_bucket_size, dim=-2),
|
143 |
+
dv.split(k_bucket_size, dim=-2),
|
144 |
+
)
|
145 |
+
|
146 |
+
for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits):
|
147 |
+
k_start_index = k_ind * k_bucket_size
|
148 |
+
|
149 |
+
attn_weights = (
|
150 |
+
torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale
|
151 |
+
)
|
152 |
+
|
153 |
+
if causal and q_start_index < (k_start_index + k_bucket_size - 1):
|
154 |
+
causal_mask = torch.ones(
|
155 |
+
(qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device
|
156 |
+
).triu(q_start_index - k_start_index + 1)
|
157 |
+
attn_weights.masked_fill_(causal_mask, max_neg_value)
|
158 |
+
|
159 |
+
exp_attn_weights = torch.exp(attn_weights - mc)
|
160 |
+
|
161 |
+
if row_mask is not None:
|
162 |
+
exp_attn_weights.masked_fill_(~row_mask, 0.0)
|
163 |
+
|
164 |
+
p = exp_attn_weights / lc
|
165 |
+
|
166 |
+
dv_chunk = torch.einsum("... i j, ... i d -> ... j d", p, doc)
|
167 |
+
dp = torch.einsum("... i d, ... j d -> ... i j", doc, vc)
|
168 |
+
|
169 |
+
D = (doc * oc).sum(dim=-1, keepdims=True)
|
170 |
+
ds = p * scale * (dp - D)
|
171 |
+
|
172 |
+
dq_chunk = torch.einsum("... i j, ... j d -> ... i d", ds, kc)
|
173 |
+
dk_chunk = torch.einsum("... i j, ... i d -> ... j d", ds, qc)
|
174 |
+
|
175 |
+
dqc.add_(dq_chunk)
|
176 |
+
dkc.add_(dk_chunk)
|
177 |
+
dvc.add_(dv_chunk)
|
178 |
+
|
179 |
+
return dq, dk, dv, None, None, None, None
|
180 |
+
|
181 |
+
|
182 |
+
class FlashAttnProcessor:
|
183 |
+
def __call__(
|
184 |
+
self,
|
185 |
+
attn: Attention,
|
186 |
+
hidden_states,
|
187 |
+
encoder_hidden_states=None,
|
188 |
+
attention_mask=None,
|
189 |
+
) -> Any:
|
190 |
+
q_bucket_size = 512
|
191 |
+
k_bucket_size = 1024
|
192 |
+
|
193 |
+
h = attn.heads
|
194 |
+
q = attn.to_q(hidden_states)
|
195 |
+
|
196 |
+
encoder_hidden_states = (
|
197 |
+
encoder_hidden_states
|
198 |
+
if encoder_hidden_states is not None
|
199 |
+
else hidden_states
|
200 |
+
)
|
201 |
+
encoder_hidden_states = encoder_hidden_states.to(hidden_states.dtype)
|
202 |
+
|
203 |
+
if hasattr(attn, "hypernetwork") and attn.hypernetwork is not None:
|
204 |
+
context_k, context_v = attn.hypernetwork.forward(
|
205 |
+
hidden_states, encoder_hidden_states
|
206 |
+
)
|
207 |
+
context_k = context_k.to(hidden_states.dtype)
|
208 |
+
context_v = context_v.to(hidden_states.dtype)
|
209 |
+
else:
|
210 |
+
context_k = encoder_hidden_states
|
211 |
+
context_v = encoder_hidden_states
|
212 |
+
|
213 |
+
k = attn.to_k(context_k)
|
214 |
+
v = attn.to_v(context_v)
|
215 |
+
del encoder_hidden_states, hidden_states
|
216 |
+
|
217 |
+
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
|
218 |
+
|
219 |
+
out = FlashAttentionFunction.apply(
|
220 |
+
q, k, v, attention_mask, False, q_bucket_size, k_bucket_size
|
221 |
+
)
|
222 |
+
|
223 |
+
out = rearrange(out, "b h n d -> b n (h d)")
|
224 |
+
|
225 |
+
out = attn.to_out[0](out)
|
226 |
+
out = attn.to_out[1](out)
|
227 |
+
return out
|