-
Notifications
You must be signed in to change notification settings - Fork 24k
/
Copy pathtest_c10d_spawn_nccl.py
200 lines (168 loc) · 7.5 KB
/
test_c10d_spawn_nccl.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
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
# Owner(s): ["oncall: distributed"]
from test_c10d_spawn import _torch_dist_nn_available, TestDistributedNNFunctions
import torch
import torch.distributed as c10d
from torch.testing._internal.common_distributed import requires_nccl, skip_if_lt_x_gpu
from torch.testing._internal.common_utils import (
run_tests,
skip_but_pass_in_sandcastle_if,
TEST_WITH_DEV_DBG_ASAN,
)
NO_NCCL = not hasattr(c10d, "ProcessGroupNCCL")
# Fails on Python-3.9, see https://github.com/pytorch/pytorch/issues/51619
# Skip dev-asan as torch + multiprocessing spawn have known issues
if not TEST_WITH_DEV_DBG_ASAN:
class TestDistributedNNFunctionsNccl(TestDistributedNNFunctions):
# Test Common Ops First.
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_broadcast(self):
self._test_broadcast("nccl")
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_reduce(self):
self._test_reduce("nccl")
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_allreduce(self):
self._test_allreduce("nccl")
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_all_gather(self):
self._test_all_gather("nccl")
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_all_to_all(self):
self._test_all_to_all("nccl")
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_all_to_all_single(self):
self._test_all_to_all_single("nccl")
# Test Ops only supported in NCCL.
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_reduce_scatter(self):
store = c10d.FileStore(self.file_name, self.world_size)
# This is required because these functions calls directly to the .dist and needs
# the world to be initialized
c10d.init_process_group(
store=store, rank=self.rank, world_size=self.world_size, backend="nccl"
)
device = torch.device(f"cuda:{self.rank}")
x0 = torch.ones(5, 5, device=device) + self.rank
x1 = torch.ones(5, 5, device=device) + self.rank + 1
x0.requires_grad = True
x1.requires_grad = True
y = torch.empty_like(x0)
expected = (
1 + self.world_size
) * self.world_size / 2 + self.world_size * self.rank
y = torch.distributed.nn.reduce_scatter(y, [x0, x1])
self.assertEqual(y, torch.ones(5, 5, device=device) * expected)
z = y.sin().sum()
z.backward()
expected_0 = (1 + self.world_size) * self.world_size / 2
expected_1 = expected_0 + self.world_size
x_s_0 = (expected_0 * torch.ones(5, 5, device=device)).cos()
x_s_1 = (expected_1 * torch.ones(5, 5, device=device)).cos()
self.assertEqual(x0.grad, x_s_0)
self.assertEqual(x1.grad, x_s_1)
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_reduce_scatter_non_contiguous(self):
store = c10d.FileStore(self.file_name, self.world_size)
# This is required because these functions calls directly to the .dist and needs
# the world to be initialized
c10d.init_process_group(
store=store, rank=self.rank, world_size=self.world_size, backend="nccl"
)
device = torch.device(f"cuda:{self.rank}")
class NonContiguousGrad(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
return input
@staticmethod
def backward(ctx, grad_output):
# Make grad non-contiguous
return grad_output.clone().transpose(0, 1)
x0 = torch.rand(5, 5, device=device, requires_grad=True)
x1 = torch.rand(5, 5, device=device, requires_grad=True)
y = torch.empty(5, 5, device=device)
y = torch.distributed.nn.reduce_scatter(y, [x0, x1])
NonContiguousGrad.apply(y).sum().backward()
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_all_reduce_non_contiguous(self):
store = c10d.FileStore(self.file_name, self.world_size)
# This is required because these functions calls directly to the .dist and needs
# the world to be initialized
c10d.init_process_group(
store=store, rank=self.rank, world_size=self.world_size, backend="nccl"
)
device = torch.device(f"cuda:{self.rank}")
class NonContiguousGrad(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
return input
@staticmethod
def backward(ctx, grad_output):
# Make grad non-contiguous
return grad_output.clone().transpose(0, 1)
x = torch.rand(5, 5, device=device, requires_grad=True)
y = torch.distributed.nn.all_reduce(x)
NonContiguousGrad.apply(y).sum().backward()
@requires_nccl()
@skip_if_lt_x_gpu(2)
@skip_but_pass_in_sandcastle_if(
not _torch_dist_nn_available, "torch.distributed.nn is not available"
)
def test_all_gather_base(self):
store = c10d.FileStore(self.file_name, self.world_size)
c10d.init_process_group(
store=store, rank=self.rank, world_size=self.world_size, backend="nccl"
)
device = torch.device(f"cuda:{self.rank}")
x = torch.ones(5, 5, device=device) + self.rank
x.requires_grad = True
output = torch.empty(5 * self.world_size, 5, device=device)
output = torch.distributed.nn.functional._all_gather_base(output, x)
self.assertEqual(output.size(), torch.Size((5 * self.world_size, 5)))
for idx in range(self.world_size):
self.assertEqual(
output[5 * idx : 5 * (idx + 1)],
torch.ones(5, 5, device=device) + idx,
)
y = torch.sum(output.view(self.world_size, 5, 5), axis=0)
z = y.sin().sum()
z.backward()
x_s = 2 * (3 * torch.ones(5, 5, device=device)).cos()
self.assertEqual(x.grad, x_s)
if __name__ == "__main__":
run_tests()