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
|
# Owner(s): ["oncall: distributed"]
import sys
import unittest
import torch
import torch.distributed as dist
import torch.distributed._functional_collectives as funcol
import torch.nn as nn
from torch.distributed._tensor import DeviceMesh, init_device_mesh, Shard
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.tensor.parallel import (
ColwiseParallel,
parallelize_module,
RowwiseParallel,
)
from torch.fx.experimental.proxy_tensor import make_fx
from torch.testing import FileCheck
from torch.testing._internal.common_utils import run_tests, TestCase
from torch.testing._internal.distributed._tensor.common_dtensor import MLPModule
from torch.testing._internal.distributed.fake_pg import FakeStore
if not dist.is_available():
print("Distributed not available, skipping tests", file=sys.stderr)
sys.exit(0)
HAS_CUDA = torch.cuda.is_available()
class TestFakePG(TestCase):
def tearDown(self):
super().tearDown()
dist.destroy_process_group()
def test_all_reduce(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=1, world_size=2, store=store)
output = torch.ones(3, 3) * dist.get_rank()
dist.all_reduce(output)
self.assertEqual(tuple(output.shape), (3, 3))
def test_allgather(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=1, world_size=2, store=store)
input_tensor = torch.ones(3, 3) * dist.get_rank()
output_tensors = [torch.empty_like(input_tensor) for _ in range(2)]
dist.all_gather(output_tensors, input_tensor)
for _, out_tensor in enumerate(output_tensors):
self.assertEqual(tuple(out_tensor.shape), (3, 3))
def test_reduce_scatter(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=1, world_size=2, store=store)
to_reduce_scatter = [torch.ones(3, 3) * rank for rank in range(2)]
output_tensor = torch.empty(3, 3)
dist.reduce_scatter(output_tensor, to_reduce_scatter)
self.assertEqual(tuple(output_tensor.shape), (3, 3))
@unittest.skipIf(not HAS_CUDA, "No CUDA")
def test_construct_fsdp(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
FSDP(nn.Linear(2, 3, device="cuda"))
@unittest.skipIf(not HAS_CUDA, "No CUDA")
def test_fsdp_fake_e2e(self):
store = dist.HashStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
my_module = nn.Sequential(
nn.Linear(2, 3, device="cuda"),
nn.ReLU(),
nn.Linear(3, 2, device="cuda"),
)
sharded_module = FSDP(my_module, use_orig_params=True)
optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001)
input = torch.randn(2, 2)
x = sharded_module(input)
loss = x.sum()
loss.backward()
optim.step()
@unittest.skipIf(not HAS_CUDA, "No CUDA")
def test_fake_pg_tracing(self):
store = dist.HashStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
default_pg = dist.distributed_c10d._get_default_group()
def allgather_fn(tensor):
return funcol.all_gather_tensor(tensor, 0, default_pg)
gm = make_fx(allgather_fn)(torch.randn(2, 2, device="cuda"))
FileCheck().check("all_gather").check("wait_tensor").run(str(gm.graph))
def test_broadcast(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
# src == rank
output = torch.ones(3, 3)
dist.broadcast(output, src=0)
self.assertEqual(tuple(output.shape), (3, 3))
# src != rank
output = torch.ones(3, 3)
dist.broadcast(output, src=1)
self.assertEqual(tuple(output.shape), (3, 3))
def test_scatter(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
# src == rank
output = torch.ones(3, 3)
to_scatter = [torch.ones(3, 3) * rank for rank in range(2)]
dist.scatter(output, to_scatter)
self.assertEqual(tuple(output.shape), (3, 3))
# src != rank
output = torch.ones(3, 3)
dist.scatter(output, None, src=1)
self.assertEqual(tuple(output.shape), (3, 3))
def test_alltoall(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
output_list = [torch.ones(3, 3) for _ in range(2)]
input_list = [torch.ones(3, 3) for _ in range(2)]
dist.all_to_all(output_list, input_list)
self.assertEqual(len(output_list), 2)
for output in output_list:
self.assertEqual(tuple(output.shape), (3, 3))
def test_alltoall_base(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
out_tensor = torch.ones(3, 3)
in_tensor = torch.ones(3, 3)
output_split = [1, 1]
input_split = [1, 1]
dist.all_to_all_single(out_tensor, in_tensor, output_split, input_split)
self.assertEqual(tuple(out_tensor.shape), (3, 3))
def test_send(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
tensor = torch.ones(3, 3)
dist.send(tensor, 1)
self.assertEqual(tuple(tensor.shape), (3, 3))
def test_recv(self):
store = FakeStore()
dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
output = torch.ones(3, 3)
dist.recv(output, 1)
self.assertEqual(tuple(output.shape), (3, 3))
@unittest.skipIf(not HAS_CUDA, "No CUDA or TP+FSDP")
def test_fsdp_tp_fake_e2e(self):
world_size = 4
tp_size = 2
store = dist.HashStore()
dist.init_process_group(
backend="fake", rank=0, world_size=world_size, store=store
)
device_mesh = DeviceMesh("cuda", torch.arange(0, world_size).view(-1, tp_size))
device_mesh = init_device_mesh(
"cuda", (world_size // tp_size, tp_size), mesh_dim_names=["dp", "tp"]
)
sequence_parallelize_plan = {
"net1": ColwiseParallel(input_layouts=Shard(0)),
"net2": RowwiseParallel(output_layouts=Shard(0)),
}
pairwise_parallelize_plan = {
"net1": ColwiseParallel(),
"net2": RowwiseParallel(),
}
for parallel_plan in [sequence_parallelize_plan, pairwise_parallelize_plan]:
my_module = parallelize_module(
MLPModule(device="cuda"),
device_mesh["tp"],
parallel_plan,
)
sharded_module = FSDP(
my_module, use_orig_params=True, device_mesh=device_mesh["dp"]
)
optim = torch.optim.Adam(sharded_module.parameters(), lr=0.0001)
for i in range(10):
dp_rank = dist.get_rank()
torch.manual_seed(i + dp_rank)
input = torch.randn(20, 10).cuda(dist.get_rank())
x = sharded_module(input)
loss = x.sum()
loss.backward()
optim.step()
if __name__ == "__main__":
run_tests()
|
