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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
|
# Owner(s): ["module: inductor"]
import math
import unittest
import torch
from torch._inductor import config
from torch._inductor.test_case import run_tests, TestCase
from torch.testing._internal.inductor_utils import HAS_CPU
def dummy_fn(x):
return torch.sigmoid(x + math.pi) / 10.0
class DummyModule(torch.nn.Module):
def forward(self, x):
return dummy_fn(x)
class TestInductorConfig(TestCase):
@classmethod
def setUpClass(cls):
super().setUpClass()
cls._saved_config = config.save_config()
def tearDown(self):
super().tearDown()
config.load_config(self._saved_config)
def test_set(self):
config.max_fusion_size = 13337
self.assertEqual(config.max_fusion_size, 13337)
self.assertEqual(config.get_config_copy()["max_fusion_size"], 13337)
config.max_fusion_size = 32
self.assertEqual(config.max_fusion_size, 32)
# a nested config
prior = config.triton.cudagraphs
config.triton.cudagraphs = not prior
self.assertEqual(config.triton.cudagraphs, not prior)
self.assertEqual(config.get_config_copy()["triton.cudagraphs"], not prior)
def test_save_load(self):
config.max_fusion_size = 123
config.triton.cudagraphs = True
saved1 = config.save_config()
config.max_fusion_size = 321
config.triton.cudagraphs = False
saved2 = config.save_config()
self.assertEqual(config.max_fusion_size, 321)
self.assertEqual(config.triton.cudagraphs, False)
config.load_config(saved1)
self.assertEqual(config.max_fusion_size, 123)
self.assertEqual(config.triton.cudagraphs, True)
config.load_config(saved2)
self.assertEqual(config.max_fusion_size, 321)
self.assertEqual(config.triton.cudagraphs, False)
def test_hasattr(self):
self.assertTrue(hasattr(config, "max_fusion_size"))
self.assertFalse(hasattr(config, "missing_name"))
def test_invalid_names(self):
self.assertRaises(AttributeError, lambda: config.does_not_exist)
self.assertRaises(AttributeError, lambda: config.triton.does_not_exist)
def store1():
config.does_not_exist = True
def store2():
config.triton.does_not_exist = True
self.assertRaises(AttributeError, store1)
self.assertRaises(AttributeError, store2)
def test_patch(self):
with config.patch(max_fusion_size=456):
self.assertEqual(config.max_fusion_size, 456)
with config.patch(max_fusion_size=789):
self.assertEqual(config.max_fusion_size, 789)
self.assertEqual(config.max_fusion_size, 456)
with config.patch({"cpp.threads": 9000, "max_fusion_size": 9001}):
self.assertEqual(config.cpp.threads, 9000)
self.assertEqual(config.max_fusion_size, 9001)
with config.patch("cpp.threads", 8999):
self.assertEqual(config.cpp.threads, 8999)
self.assertEqual(config.cpp.threads, 9000)
@unittest.skipIf(not HAS_CPU, "requires C++ compiler")
def test_compile_api(self):
# these are mostly checking config processing doesn't blow up with exceptions
x = torch.randn(8)
y = dummy_fn(x)
checks = [
{},
{"mode": "default"},
{"mode": "reduce-overhead"},
{"mode": "max-autotune"},
{
"options": {
"max-fusion-size": 128,
"unroll_reductions_threshold": 32,
"triton.cudagraphs": False,
}
},
{"dynamic": True},
{"fullgraph": True, "backend": "inductor"},
{"disable": True},
]
for kwargs in checks:
torch._dynamo.reset()
opt_fn = torch.compile(dummy_fn, **kwargs)
torch.testing.assert_allclose(
opt_fn(x), y, msg=f"torch.compile(..., **{kwargs!r}) failed"
)
def test_get_compiler_config(self):
from torch._inductor import config as inductor_default_config
default_cudagraphs = inductor_default_config.triton.cudagraphs
# nn.Module: should update default config with a new value
model = DummyModule()
optimized_module = torch.compile(
model, options={"triton.cudagraphs": not default_cudagraphs}
)
compiler_config = optimized_module.get_compiler_config()
self.assertEqual(compiler_config["triton.cudagraphs"], not default_cudagraphs)
# nn.Module: keep default config
model = DummyModule()
optimized_module = torch.compile(model)
compiler_config = optimized_module.get_compiler_config()
self.assertEqual(
compiler_config["triton.cudagraphs"],
default_cudagraphs,
)
# compile user func: should update default config with a new value
optimized_module = torch.compile(
dummy_fn, options={"triton.cudagraphs": not default_cudagraphs}
)
compiler_config = optimized_module.get_compiler_config()
self.assertEqual(compiler_config["triton.cudagraphs"], not default_cudagraphs)
# compile user func: keep default config
optimized_module = torch.compile(dummy_fn)
compiler_config = optimized_module.get_compiler_config()
self.assertEqual(
compiler_config["triton.cudagraphs"],
default_cudagraphs,
)
# backend=eager: expect None
optimized_module = torch.compile(dummy_fn, backend="eager")
compiler_config = optimized_module.get_compiler_config()
self.assertTrue(compiler_config is None)
def test_compile_api_passes_config(self):
# ensure configs are actually passed down to inductor
self.assertRaises(
torch._dynamo.exc.BackendCompilerFailed,
lambda: torch.compile(dummy_fn, options={"_raise_error_for_testing": True})(
torch.randn(10)
),
)
def test_api_options(self):
reduce_overhead_opts = torch._inductor.list_mode_options("reduce-overhead")
self.assertEqual(reduce_overhead_opts["triton.cudagraphs"], True)
self.assertEqual(reduce_overhead_opts.get("max_autotune", False), False)
max_autotune_opts = torch._inductor.list_mode_options("max-autotune")
self.assertEqual(max_autotune_opts["max_autotune"], True)
self.assertEqual(max_autotune_opts["triton.cudagraphs"], True)
max_autotune_opts = torch._inductor.list_mode_options(
"max-autotune", dynamic=True
)
self.assertEqual(max_autotune_opts["max_autotune"], True)
self.assertEqual(max_autotune_opts["triton.cudagraphs"], True)
max_autotune_no_cudagraphs_opts = torch._inductor.list_mode_options(
"max-autotune-no-cudagraphs"
)
self.assertEqual(max_autotune_no_cudagraphs_opts["max_autotune"], True)
self.assertEqual(
max_autotune_no_cudagraphs_opts.get("triton.cudagraphs", False), False
)
def test_invalid_backend(self):
self.assertRaises(
torch._dynamo.exc.InvalidBackend,
lambda: torch.compile(dummy_fn, backend="does_not_exist")(torch.randn(10)),
)
def test_non_inductor_backend(self):
def assert_options(expected_mode=None, expected_options=None):
def backend(gm, _, *, mode=None, options=None):
nonlocal call_count
self.assertEqual(mode, expected_mode)
self.assertEqual(options, expected_options)
call_count += 1
return gm
return backend
inp = torch.randn(8)
def fn(x):
return x + 1
for mode, options in [
(None, None),
("fast-mode", None),
(None, {"foo": "bar"}),
]:
call_count = 0
torch.compile(
fn, backend=assert_options(mode, options), mode=mode, options=options
)(inp)
torch._dynamo.reset()
self.assertEqual(call_count, 1)
if __name__ == "__main__":
run_tests()
|
