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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
|
# Owner(s): ["module: dynamo"]
# flake8: noqa
import functools
import torch
import torch._dynamo.test_case
import torch._dynamo.testing
import torch._dynamo.utils
from torch import _inductor as inductor
from torch._dynamo import compiled_autograd
from torch._dynamo._trace_wrapped_higher_order_op import trace_wrapped
from torch._dynamo.testing import normalize_gm
from torch._dynamo.utils import counters
from torch.fx.experimental.proxy_tensor import make_fx
def _multiply(x):
return x * x
def _multiply_invoke(grad):
return trace_wrapped(grad, fn=_multiply)
class BackwardHigherOrderOpTests(torch._dynamo.test_case.TestCase):
def test_invoke_in_eager(self):
x = torch.tensor([0.5, 0.5], requires_grad=True)
y = torch.tensor([0.5, 0.5], requires_grad=True)
def fn(x, y):
x.register_hook(_multiply_invoke)
return x * y
out = fn(x, y)
grad_out = torch.tensor([2.0, 2.0])
out.backward(grad_out)
self.assertEqual(x.grad, y * grad_out)
def test_invoke_in_pt2(self):
for backend in ["eager", "aot_eager", "inductor"]:
torch._dynamo.reset()
x = torch.tensor([0.5, 0.5], requires_grad=True)
y = torch.tensor([0.5, 0.5], requires_grad=True)
def fn(x, y):
x.register_hook(_multiply_invoke)
return x * y
fn = torch.compile(fn, backend=backend)
out = fn(x, y)
grad_out = torch.tensor([2.0, 2.0])
out.backward(grad_out)
self.assertEqual(x.grad, grad_out * y)
def test_invoke_make_fx_forward_contrived(self):
x = torch.tensor([0.5, 0.5], requires_grad=True)
out = make_fx(_multiply_invoke)(x)
self.assertEqual(out(x), torch.tensor([0.25, 0.25]))
actual = normalize_gm(out.print_readable(False))
self.assertExpectedInline(
actual,
"""\
class _multiply_invoke(torch.nn.Module):
def forward(self, grad_1: "f32[2]"):
trace_wrapped: "f32[2]" = torch__dynamo__trace_wrapped_higher_order_op_self_invoke(grad_1); grad_1 = None
return trace_wrapped
""",
)
def test_invoke_make_bw(self):
x = torch.tensor([0.5, 0.5], requires_grad=True)
def fwd(x):
z = x * x
return z + z
res = fwd(x)
res.backward(torch.tensor([1.0, 1.0]))
out = make_fx(_multiply_invoke)(x.grad)
self.assertEqual(out(x.grad), torch.tensor([4.0, 4.0]))
actual = normalize_gm(out.print_readable(False))
self.assertExpectedInline(
actual,
"""\
class _multiply_invoke(torch.nn.Module):
def forward(self, grad_1: "f32[2]"):
trace_wrapped: "f32[2]" = torch__dynamo__trace_wrapped_higher_order_op_self_invoke(grad_1); grad_1 = None
return trace_wrapped
""",
)
def test_invoke_in_pt2_compiled_autograd(self):
graph = None
def compiler_fn(gm):
def inner_compiler(gm_, example_inputs_):
nonlocal graph
self.assertEqual(graph, None)
graph = gm_
return inductor.compile(gm_, example_inputs_)
return torch.compile(
gm, backend=inner_compiler, fullgraph=True, dynamic=True
)
for backend in ["eager", "aot_eager", "inductor"]:
torch._dynamo.reset()
x = torch.tensor([0.5, 0.5], requires_grad=True)
y = torch.tensor([0.5, 0.5], requires_grad=True)
def fn(x, y):
x.register_hook(_multiply_invoke)
return x + y
fn = torch.compile(fn, backend=backend)
out = fn(x, y)
grad_out = torch.tensor([2.0, 2.0])
with compiled_autograd._enable(compiler_fn):
out.backward(grad_out)
actual = normalize_gm(graph.print_readable(False))
self.assertEqual(x.grad, grad_out * grad_out)
self.assertExpectedInline(
actual,
"""\
class GraphModule(torch.nn.Module):
def forward(self, L_inputs_ : list):
l_inputs_ = L_inputs_
getitem: "f32[s0]" = l_inputs_[0]; l_inputs_ = None
new_grad: "f32[s0]" = torch.clone(getitem)
result: "f32[s0]" = getitem * getitem; getitem = None
new_grad_1: "f32[s0]" = torch.clone(result); result = None
return (new_grad, new_grad_1)
""",
)
graph = None
def test_invoke_in_pt2_compiled_autograd_side_effect(self):
def _side_effect_stateful_fn2(x, obj):
obj.counter = obj.counter + 1
return _multiply(x)
def _side_effectful_invoke2(grad, fn):
return trace_wrapped(grad, fn=fn)
graph = None
def compiler_fn(gm):
def inner_compiler(gm_, example_inputs_):
nonlocal graph
self.assertEqual(graph, None)
graph = gm_
return inductor.compile(gm_, example_inputs_)
return torch.compile(
gm, backend=inner_compiler, fullgraph=True, dynamic=True
)
for backend in ["eager", "aot_eager", "inductor"]:
torch._dynamo.reset()
x = torch.tensor([0.5, 0.5], requires_grad=True)
y = torch.tensor([0.5, 0.5], requires_grad=True)
class MyObj:
def __init__(self) -> None:
self.counter = 0
obj = MyObj()
inner_fn = functools.partial(_side_effect_stateful_fn2, obj=obj)
hook_fn = functools.partial(_side_effectful_invoke2, fn=inner_fn)
x.register_hook(hook_fn)
def fn(x, y):
return x + y
fn = torch.compile(fn, backend=backend, fullgraph=True)
out = fn(x, y)
grad_out = torch.tensor([2.0, 2.0])
with compiled_autograd._enable(compiler_fn):
out.backward(grad_out)
actual = normalize_gm(graph.print_readable(False))
self.assertEqual(obj.counter, 1)
self.assertEqual(x.grad, grad_out + grad_out)
self.assertExpectedInline(
actual,
"""\
class GraphModule(torch.nn.Module):
def forward(self, L_inputs_ : list, L_hooks_0_keywords_fn_keywords_obj_counter: "Sym(s1)"):
l_inputs_ = L_inputs_
l_hooks_0_keywords_fn_keywords_obj_counter = L_hooks_0_keywords_fn_keywords_obj_counter
getitem: "f32[s0]" = l_inputs_[0]; l_inputs_ = None
new_grad: "f32[s0]" = torch.clone(getitem)
add: "Sym(s1 + 1)" = l_hooks_0_keywords_fn_keywords_obj_counter + 1; l_hooks_0_keywords_fn_keywords_obj_counter = None
result: "f32[s0]" = getitem * getitem; getitem = None
new_grad_1: "f32[s0]" = torch.clone(result); result = None
return (new_grad, new_grad_1, add)
""",
)
out = fn(x, y)
out.backward(grad_out)
self.assertEqual(obj.counter, 2)
out = fn(x, y)
out.backward(grad_out)
self.assertEqual(obj.counter, 3)
graph = None
def test_invoke_in_pt2_compiled_autograd_graph_breaks(self):
def _graph_breaking_fn(x):
print("Boo!")
return _multiply(x)
def _graph_break_invoke(grad):
return trace_wrapped(grad, fn=_graph_breaking_fn)
def compiler_fn(gm):
return torch.compile(gm, backend="inductor", fullgraph=True, dynamic=True)
for backend in ["eager", "aot_eager", "inductor"]:
torch._dynamo.reset()
x = torch.tensor([0.5, 0.5], requires_grad=True)
y = torch.tensor([0.5, 0.5], requires_grad=True)
def fn(x, y):
x.register_hook(_graph_break_invoke)
return x + y
fn = torch.compile(fn, backend=backend, fullgraph=True)
out = fn(x, y)
grad_out = torch.tensor([2.0, 2.0])
with self.assertRaisesRegex(
torch._dynamo.exc.Unsupported,
"print",
):
with compiled_autograd._enable(compiler_fn):
out.backward(grad_out)
graph = None
if __name__ == "__main__":
from torch._dynamo.test_case import run_tests
run_tests()
|
