Spaces:
Running
on
Zero
Running
on
Zero
Upload folder using huggingface_hub
Browse files- modules/Attention/AttentionMethods.py +113 -36
- modules/BlackForest/Flux.py +1 -1
modules/Attention/AttentionMethods.py
CHANGED
@@ -4,9 +4,17 @@ except ImportError:
|
|
4 |
pass
|
5 |
import torch
|
6 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
def attention_xformers(
|
9 |
-
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask=None, skip_reshape=False
|
10 |
) -> torch.Tensor:
|
11 |
"""#### Make an attention call using xformers. Fastest attention implementation.
|
12 |
|
@@ -20,31 +28,84 @@ def attention_xformers(
|
|
20 |
#### Returns:
|
21 |
- `torch.Tensor`: The output tensor.
|
22 |
"""
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
out
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
|
46 |
def attention_pytorch(
|
47 |
-
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask=None, skip_reshape=False
|
48 |
) -> torch.Tensor:
|
49 |
"""#### Make an attention call using PyTorch.
|
50 |
|
@@ -58,19 +119,35 @@ def attention_pytorch(
|
|
58 |
#### Returns:
|
59 |
- `torch.Tensor`: The output tensor.
|
60 |
"""
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
def xformers_attention(
|
76 |
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
|
|
|
4 |
pass
|
5 |
import torch
|
6 |
|
7 |
+
BROKEN_XFORMERS = False
|
8 |
+
try:
|
9 |
+
x_vers = xformers.__version__
|
10 |
+
# XFormers bug confirmed on all versions from 0.0.21 to 0.0.26 (q with bs bigger than 65535 gives CUDA error)
|
11 |
+
BROKEN_XFORMERS = x_vers.startswith("0.0.2") and not x_vers.startswith("0.0.20")
|
12 |
+
except:
|
13 |
+
pass
|
14 |
+
|
15 |
|
16 |
def attention_xformers(
|
17 |
+
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask=None, skip_reshape=False, flux=False
|
18 |
) -> torch.Tensor:
|
19 |
"""#### Make an attention call using xformers. Fastest attention implementation.
|
20 |
|
|
|
28 |
#### Returns:
|
29 |
- `torch.Tensor`: The output tensor.
|
30 |
"""
|
31 |
+
if not flux:
|
32 |
+
b, _, dim_head = q.shape
|
33 |
+
dim_head //= heads
|
34 |
+
|
35 |
+
q, k, v = map(
|
36 |
+
lambda t: t.unsqueeze(3)
|
37 |
+
.reshape(b, -1, heads, dim_head)
|
38 |
+
.permute(0, 2, 1, 3)
|
39 |
+
.reshape(b * heads, -1, dim_head)
|
40 |
+
.contiguous(),
|
41 |
+
(q, k, v),
|
42 |
+
)
|
43 |
+
|
44 |
+
out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=mask)
|
45 |
+
|
46 |
+
out = (
|
47 |
+
out.unsqueeze(0)
|
48 |
+
.reshape(b, heads, -1, dim_head)
|
49 |
+
.permute(0, 2, 1, 3)
|
50 |
+
.reshape(b, -1, heads * dim_head)
|
51 |
+
)
|
52 |
+
return out
|
53 |
+
else:
|
54 |
+
if skip_reshape:
|
55 |
+
b, _, _, dim_head = q.shape
|
56 |
+
else:
|
57 |
+
b, _, dim_head = q.shape
|
58 |
+
dim_head //= heads
|
59 |
+
|
60 |
+
disabled_xformers = False
|
61 |
+
|
62 |
+
if BROKEN_XFORMERS:
|
63 |
+
if b * heads > 65535:
|
64 |
+
disabled_xformers = True
|
65 |
+
|
66 |
+
if not disabled_xformers:
|
67 |
+
if torch.jit.is_tracing() or torch.jit.is_scripting():
|
68 |
+
disabled_xformers = True
|
69 |
+
|
70 |
+
if disabled_xformers:
|
71 |
+
return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape)
|
72 |
+
|
73 |
+
if skip_reshape:
|
74 |
+
q, k, v = map(
|
75 |
+
lambda t: t.reshape(b * heads, -1, dim_head),
|
76 |
+
(q, k, v),
|
77 |
+
)
|
78 |
+
else:
|
79 |
+
q, k, v = map(
|
80 |
+
lambda t: t.reshape(b, -1, heads, dim_head),
|
81 |
+
(q, k, v),
|
82 |
+
)
|
83 |
+
|
84 |
+
if mask is not None:
|
85 |
+
pad = 8 - q.shape[1] % 8
|
86 |
+
mask_out = torch.empty(
|
87 |
+
[q.shape[0], q.shape[1], q.shape[1] + pad], dtype=q.dtype, device=q.device
|
88 |
+
)
|
89 |
+
mask_out[:, :, : mask.shape[-1]] = mask
|
90 |
+
mask = mask_out[:, :, : mask.shape[-1]]
|
91 |
+
|
92 |
+
out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=mask)
|
93 |
+
|
94 |
+
if skip_reshape:
|
95 |
+
out = (
|
96 |
+
out.unsqueeze(0)
|
97 |
+
.reshape(b, heads, -1, dim_head)
|
98 |
+
.permute(0, 2, 1, 3)
|
99 |
+
.reshape(b, -1, heads * dim_head)
|
100 |
+
)
|
101 |
+
else:
|
102 |
+
out = out.reshape(b, -1, heads * dim_head)
|
103 |
+
|
104 |
+
return out
|
105 |
|
106 |
|
107 |
def attention_pytorch(
|
108 |
+
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, heads: int, mask=None, skip_reshape=False, flux=False
|
109 |
) -> torch.Tensor:
|
110 |
"""#### Make an attention call using PyTorch.
|
111 |
|
|
|
119 |
#### Returns:
|
120 |
- `torch.Tensor`: The output tensor.
|
121 |
"""
|
122 |
+
if not flux:
|
123 |
+
b, _, dim_head = q.shape
|
124 |
+
dim_head //= heads
|
125 |
+
q, k, v = map(
|
126 |
+
lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
|
127 |
+
(q, k, v),
|
128 |
+
)
|
129 |
+
|
130 |
+
out = torch.nn.functional.scaled_dot_product_attention(
|
131 |
+
q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False
|
132 |
+
)
|
133 |
+
out = out.transpose(1, 2).reshape(b, -1, heads * dim_head)
|
134 |
+
return out
|
135 |
+
else:
|
136 |
+
if skip_reshape:
|
137 |
+
b, _, _, dim_head = q.shape
|
138 |
+
else:
|
139 |
+
b, _, dim_head = q.shape
|
140 |
+
dim_head //= heads
|
141 |
+
q, k, v = map(
|
142 |
+
lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
|
143 |
+
(q, k, v),
|
144 |
+
)
|
145 |
+
|
146 |
+
out = torch.nn.functional.scaled_dot_product_attention(
|
147 |
+
q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False
|
148 |
+
)
|
149 |
+
out = out.transpose(1, 2).reshape(b, -1, heads * dim_head)
|
150 |
+
return out
|
151 |
|
152 |
def xformers_attention(
|
153 |
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor
|
modules/BlackForest/Flux.py
CHANGED
@@ -29,7 +29,7 @@ def attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, pe: torch.Tenso
|
|
29 |
"""
|
30 |
q, k = apply_rope(q, k, pe)
|
31 |
heads = q.shape[1]
|
32 |
-
x = Attention.optimized_attention(q, k, v, heads, skip_reshape=True)
|
33 |
return x
|
34 |
|
35 |
# Define the rotary positional encoding (RoPE)
|
|
|
29 |
"""
|
30 |
q, k = apply_rope(q, k, pe)
|
31 |
heads = q.shape[1]
|
32 |
+
x = Attention.optimized_attention(q, k, v, heads, skip_reshape=True, flux=True)
|
33 |
return x
|
34 |
|
35 |
# Define the rotary positional encoding (RoPE)
|