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
|
# Owner(s): ["module: inductor"]
import copy
import os
import shutil
import tempfile
import types
import torch
import torch._export
import torch._inductor
import torch.export._trace
import torch.fx._pytree as fx_pytree
from torch._dynamo.testing import same
from torch._inductor import config
from torch._inductor.test_case import TestCase
from torch.testing import FileCheck
from torch.testing._internal.common_utils import IS_FBCODE
from torch.utils import _pytree as pytree
class WrapperModule(torch.nn.Module):
def __init__(self, model):
super().__init__()
self.model = model
def forward(self, *args, **kwargs):
return self.model(*args, **kwargs)
class AOTIRunnerUtil:
@staticmethod
def compile(
model,
example_inputs,
options=None,
dynamic_shapes=None,
disable_constraint_solver=False,
):
if not isinstance(model, torch.nn.Module):
model = WrapperModule(model)
# The exact API is subject to change
if torch._inductor.config.is_predispatch:
ep = torch.export._trace._export(
model, example_inputs, dynamic_shapes=dynamic_shapes, pre_dispatch=True
)
gm = ep.module()
else:
gm = torch.export._trace._export_to_torch_ir(
model,
example_inputs,
dynamic_shapes=dynamic_shapes,
disable_constraint_solver=disable_constraint_solver,
# Disabling this flag, because instead we can rely on the mapping
# dynamo_flat_name_to_original_fqn which is coming from Dynamo.
restore_fqn=False,
)
if IS_FBCODE:
from deeplearning.aot_inductor.extern_node_thrift_serializer import (
thrift_serializer,
)
if options is None:
options = {}
options["extern_node_serializer"] = thrift_serializer
with torch.no_grad():
so_path = torch._inductor.aot_compile(gm, example_inputs, options=options) # type: ignore[arg-type]
return so_path
@staticmethod
def load_runner(device, so_path):
if IS_FBCODE:
from .fb import test_aot_inductor_model_runner_pybind # @manual
with tempfile.TemporaryDirectory() as temp_dir:
# copy *.so file to a unique path just before loading
# to avoid stale dlopen handles when an updated *.so
# from the same path is loaded repetitively in a test
temp_so_path = os.path.join(temp_dir, "model.so")
shutil.copy(so_path, temp_so_path)
# We also need to copy over the serialized extern_kernel_nodes for custom ops
extern_kernel_nodes_path = f"{so_path[:-3]}.json"
if os.path.isfile(extern_kernel_nodes_path):
temp_extern_kernel_nodes_path = os.path.join(temp_dir, "model.json")
shutil.copy(extern_kernel_nodes_path, temp_extern_kernel_nodes_path)
return test_aot_inductor_model_runner_pybind.Runner(
temp_so_path, device == "cpu"
)
else:
if device == "cpu":
return torch._C._aoti.AOTIModelContainerRunnerCpu(so_path, 1)
elif device == "xpu":
return torch._C._aoti.AOTIModelContainerRunnerXpu(so_path, 1, device)
else:
return torch._C._aoti.AOTIModelContainerRunnerCuda(so_path, 1, device)
@staticmethod
def load(device, so_path):
# TODO: unify fbcode and oss behavior to only use torch._export.aot_load
if IS_FBCODE:
runner = AOTIRunnerUtil.load_runner(device, so_path)
def optimized(*args, **kwargs):
call_spec = runner.get_call_spec()
in_spec = pytree.treespec_loads(call_spec[0])
out_spec = pytree.treespec_loads(call_spec[1])
flat_inputs = fx_pytree.tree_flatten_spec((args, kwargs), in_spec)
flat_inputs = [x for x in flat_inputs if isinstance(x, torch.Tensor)]
flat_outputs = runner.run(flat_inputs)
return pytree.tree_unflatten(flat_outputs, out_spec)
return optimized
else:
return torch._export.aot_load(so_path, device)
@staticmethod
def run(
device,
model,
example_inputs,
options=None,
dynamic_shapes=None,
disable_constraint_solver=False,
):
so_path = AOTIRunnerUtil.compile(
model,
example_inputs,
options=options,
dynamic_shapes=dynamic_shapes,
disable_constraint_solver=disable_constraint_solver,
)
optimized = AOTIRunnerUtil.load(device, so_path)
return optimized(*example_inputs)
@staticmethod
def run_multiple(
device,
model,
list_example_inputs,
options=None,
dynamic_shapes=None,
):
so_path = AOTIRunnerUtil.compile(
model,
list_example_inputs[0],
options=options,
dynamic_shapes=dynamic_shapes,
)
optimized = AOTIRunnerUtil.load(device, so_path)
list_output_tensors = []
for example_inputs in list_example_inputs:
list_output_tensors.append(optimized(*example_inputs))
return list_output_tensors
def check_model(
self: TestCase,
model,
example_inputs,
options=None,
dynamic_shapes=None,
disable_constraint_solver=False,
atol=None,
rtol=None,
):
with torch.no_grad(), config.patch(
{
"aot_inductor.allow_stack_allocation": self.allow_stack_allocation,
"aot_inductor.use_minimal_arrayref_interface": self.use_minimal_arrayref_interface,
}
):
torch.manual_seed(0)
if not isinstance(model, types.FunctionType):
model = model.to(self.device)
ref_model = copy.deepcopy(model)
ref_inputs = copy.deepcopy(example_inputs)
expected = ref_model(*ref_inputs)
torch.manual_seed(0)
actual = AOTIRunnerUtil.run(
self.device,
model,
example_inputs,
options,
dynamic_shapes,
disable_constraint_solver,
)
self.assertEqual(actual, expected, atol=atol, rtol=rtol)
def check_model_with_multiple_inputs(
self: TestCase,
model,
list_example_inputs,
options=None,
dynamic_shapes=None,
):
with torch.no_grad(), config.patch(
{
"aot_inductor.allow_stack_allocation": self.allow_stack_allocation,
"aot_inductor.use_minimal_arrayref_interface": self.use_minimal_arrayref_interface,
}
):
torch.manual_seed(0)
model = model.to(self.device)
ref_model = copy.deepcopy(model)
ref_inputs = copy.deepcopy(list_example_inputs)
list_expected = [ref_model(*inputs) for inputs in ref_inputs]
torch.manual_seed(0)
list_actual = AOTIRunnerUtil.run_multiple(
self.device, model, list_example_inputs, options, dynamic_shapes
)
self.assertTrue(same(list_actual, list_expected))
def code_check_count(
self: TestCase,
model,
example_inputs,
target_str: str,
target_count: int,
):
with torch.no_grad(), config.patch(
{
"aot_inductor.allow_stack_allocation": self.allow_stack_allocation,
"aot_inductor.use_minimal_arrayref_interface": self.use_minimal_arrayref_interface,
}
):
so_path = torch._export.aot_compile(model, example_inputs)
with open(os.path.splitext(so_path)[0] + ".cpp") as cpp:
src_code = cpp.read()
FileCheck().check_count(
target_str,
target_count,
exactly=True,
).run(src_code)
|
