/
test_sparse_fused_attention_op.py
163 lines (133 loc) · 5.3 KB
/
test_sparse_fused_attention_op.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import math
import re
import copy
import unittest
import numpy as np
import paddle
import paddle.fluid.core as core
from paddle.fluid.framework import _test_eager_guard
def get_cuda_version():
result = os.popen("nvcc --version").read()
regex = r'release (\S+),'
match = re.search(regex, result)
if match:
num = str(match.group(1))
integer, decimal = num.split('.')
return int(integer) * 1000 + int(float(decimal) * 10)
else:
return -1
@unittest.skipIf(
not core.is_compiled_with_cuda() or get_cuda_version() < 11070,
"core is not compiled with CUDA and cuda version need larger than or equal to 11.7"
)
class TestSparseAttentionAPI1(unittest.TestCase):
def setUp(self):
self.batch_size = 16
self.num_heads = 16
self.seq_len = 128
self.head_dim = 16
self.dtype = 'float64'
self.use_mask = True
def test_dygraph(self):
with _test_eager_guard():
self.shape = [
self.batch_size, self.num_heads, self.seq_len, self.head_dim
]
query = paddle.rand(self.shape, self.dtype)
key = paddle.rand(self.shape, self.dtype)
value = paddle.rand(self.shape, self.dtype)
query.stop_gradient = False
key.stop_gradient = False
value.stop_gradient = False
mask = paddle.nn.functional.dropout(paddle.ones(
[self.seq_len, self.seq_len]),
mode='downscale_in_infer')
mask = mask.expand(
[self.batch_size, self.num_heads, self.seq_len, self.seq_len])
sp_mask = mask.reshape([-1, self.seq_len,
self.seq_len]).to_sparse_csr()
query_sp = copy.deepcopy(query)
key_sp = copy.deepcopy(key)
value_sp = copy.deepcopy(value)
query_sp.stop_gradient = False
key_sp.stop_gradient = False
value_sp.stop_gradient = False
if self.use_mask:
kp_mask = paddle.randint(
0, 2, [self.batch_size, self.seq_len]).astype(self.dtype)
attn_mask = paddle.randint(
0, 2, [self.seq_len, self.seq_len]).astype(self.dtype)
sdd = paddle.matmul(query, key, False, True) / math.sqrt(
float(self.head_dim))
sdd = sdd + (
(mask * kp_mask.unsqueeze([1, 2]) * attn_mask) - 1.0) * 1e9
softmax = paddle.nn.functional.softmax(sdd)
output = paddle.matmul(softmax, value)
output.backward()
output_sp = paddle.incubate.sparse.nn.functional.attention(
query_sp, key_sp, value_sp, sp_mask, kp_mask, attn_mask)
output_sp.backward()
else:
sdd = paddle.matmul(query, key, False, True) / math.sqrt(
float(self.head_dim))
sdd = sdd + (mask - 1.0) * 1e9
softmax = paddle.nn.functional.softmax(sdd)
output = paddle.matmul(softmax, value)
output.backward()
output_sp = paddle.incubate.sparse.nn.functional.attention(
query_sp, key_sp, value_sp, sp_mask)
output_sp.backward()
self.assertTrue(np.allclose(output_sp.numpy(), output.numpy()))
self.assertTrue(
np.allclose(query_sp.grad.numpy(), query.grad.numpy()))
self.assertTrue(np.allclose(key_sp.grad.numpy(), key.grad.numpy()))
self.assertTrue(
np.allclose(value_sp.grad.numpy(), value.grad.numpy()))
class TestSparseAttentionAPI2(TestSparseAttentionAPI1):
def setUp(self):
self.batch_size = 16
self.num_heads = 16
self.seq_len = 128
self.head_dim = 32
self.dtype = 'float64'
self.use_mask = False
class TestSparseAttentionAPI3(TestSparseAttentionAPI1):
def setUp(self):
self.batch_size = 16
self.num_heads = 16
self.seq_len = 512
self.head_dim = 16
self.dtype = 'float64'
self.use_mask = True
class TestSparseAttentionAPI4(TestSparseAttentionAPI1):
def setUp(self):
self.batch_size = 16
self.num_heads = 16
self.seq_len = 512
self.head_dim = 32
self.dtype = 'float64'
self.use_mask = False
class TestSparseAttentionAPI5(TestSparseAttentionAPI1):
def setUp(self):
self.batch_size = 16
self.num_heads = 16
self.seq_len = 512
self.head_dim = 64
self.dtype = 'float64'
self.use_mask = True
if __name__ == '__main__':
unittest.main()