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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
|
# Owner(s): ["module: cuda graphs"]
import functools
import unittest
import torch
import torch._dynamo
import torch._dynamo.config
import torch._dynamo.test_case
import torch._dynamo.testing
from torch._dynamo.testing import same
from torch.testing._internal.common_utils import TEST_CUDA_GRAPH
def composed(*decs):
def deco(f):
for dec in reversed(decs):
f = dec(f)
return f
return deco
def assert_aot_autograd_counter(ok=True):
def deco(f):
@functools.wraps(f)
def wrap(self, *args, **kwargs):
torch._dynamo.utils.counters.clear()
r = f(self, *args, **kwargs)
c_ok = torch._dynamo.utils.counters["aot_autograd"]["ok"]
c_not_ok = torch._dynamo.utils.counters["aot_autograd"]["not_ok"]
if ok:
self.assertGreater(c_ok, 0)
self.assertEqual(c_not_ok, 0)
else:
self.assertEqual(c_ok, 0)
self.assertGreater(c_not_ok, 0)
return r
return wrap
return deco
def patch_all(ok=True):
return composed(
torch._dynamo.config.patch(
verify_correctness=True, automatic_dynamic_shapes=True
),
assert_aot_autograd_counter(ok),
)
N_ITERS = 5
@unittest.skipIf(not torch.cuda.is_available(), "these tests require cuda")
class TestAotCudagraphs(torch._dynamo.test_case.TestCase):
@patch_all()
def test_basic(self):
def model(x, y):
return (x + y) * y
@torch.compile(backend="cudagraphs")
def fn(x, y):
for i in range(N_ITERS):
loss = model(x, y).sum()
loss.backward()
x = torch.randn(3, device="cuda", requires_grad=True)
y = torch.randn(3, device="cuda")
fn(x, y)
@patch_all()
def test_dtoh(self):
def model(x, y):
a = x + y
b = a.cpu() * 3
return b
@torch.compile(backend="cudagraphs")
def fn(x, y):
for i in range(N_ITERS):
loss = model(x, y).sum()
loss.backward()
x = torch.randn(3, device="cuda", requires_grad=True)
y = torch.randn(3, device="cuda")
fn(x, y)
@patch_all()
def test_htod(self):
def model(x, y):
a = x + y
return a * 3
@torch.compile(backend="cudagraphs")
def fn(x, y):
for i in range(N_ITERS):
loss = model(x, y).sum()
loss.backward()
x = torch.randn(3, device="cuda", requires_grad=True)
y = torch.randn((), device="cpu")
fn(x, y)
def test_mutate_input(self):
def model(x, y):
y.add_(3)
return x * y
@torch.compile(backend="cudagraphs")
def fn(x, y):
for i in range(N_ITERS):
with self.subTest(i):
y_orig = y.clone()
loss = model(x, y).sum()
self.assertTrue(same(y, y_orig + 3))
loss.backward()
x = torch.randn(3, device="cuda", requires_grad=True)
y = torch.randn(3, device="cuda")
fn(x, y)
@patch_all()
def test_mutate_constant(self):
def model(x, y):
c = torch.tensor(1)
c.add_(2)
return x * y * 0 + c
@torch.compile(backend="cudagraphs")
def fn(x, y):
for i in range(N_ITERS):
with self.subTest(i):
loss = model(x, y).sum()
self.assertTrue(same(loss, torch.tensor(3.0, device="cuda")))
loss.backward()
x = torch.randn(1, device="cuda", requires_grad=True)
y = torch.randn(1, device="cuda")
fn(x, y)
@patch_all()
def test_factory(self):
def model(y):
x = torch.zeros(3, device="cuda:0")
x.add_(3)
return x * y
@torch.compile(backend="cudagraphs")
def fn(y):
for i in range(N_ITERS):
with self.subTest(i):
loss = model(y).sum()
loss.backward()
y = torch.randn(3, device="cuda:0", requires_grad=True)
fn(y)
@patch_all()
def test_mutated_metadata(self):
# more tortured example at
# https://github.com/pytorch/pytorch/issues/81385
def model(x):
x = x.clone()
x.resize_(20)
x.fill_(2)
return x
@torch.compile(backend="cudagraphs")
def fn(x):
for i in range(N_ITERS):
with self.subTest(i):
rx = model(x)
self.assertTrue(same(rx, torch.full((20,), 2.0, device="cuda:0")))
x = torch.empty(0, device="cuda:0")
fn(x)
@patch_all()
def test_dead_fill(self):
def model(x):
x = x.clone()
y = x[0:0]
x.fill_(2)
y.fill_(3)
return x, y
@torch.compile(backend="cudagraphs")
def fn(x):
for i in range(N_ITERS):
with self.subTest(i):
rx, ry = model(x)
self.assertTrue(same(rx, torch.full((20,), 2.0, device="cuda:0")))
self.assertTrue(same(ry, torch.empty(0, device="cuda:0")))
x = torch.empty(20, device="cuda:0")
fn(x)
if __name__ == "__main__":
from torch._dynamo.test_case import run_tests
if not TEST_CUDA_GRAPH:
if __name__ == "__main__":
import sys
sys.exit(0)
raise unittest.SkipTest("cuda graph test is skipped")
run_tests()
|
