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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
|
# Owner(s): ["oncall: distributed"]
import torch
import torch.distributed.checkpoint as dist_cp
from torch.distributed._tensor import (
distribute_tensor,
init_device_mesh,
Replicate,
Shard,
zeros,
)
from torch.testing._internal.common_utils import run_tests
from torch.testing._internal.distributed._tensor.common_dtensor import (
DTensorTestBase,
skip_if_lt_x_gpu,
with_comms,
)
from torch.testing._internal.distributed.checkpoint_utils import with_temp_dir
CHECKPOINT_DIR = "checkpoint"
ONE_D_PLACEMENTS = [
[Shard(0)],
[Replicate()],
]
ONE_D_TO_ONE_D_PLACEMENTS = [
([Replicate()], [Shard(0)]),
([Shard(0)], [Replicate()]),
]
TWO_D_PLACEMENTS = [
[Replicate(), Replicate()],
[Replicate(), Shard(0)],
[Shard(0), Replicate()],
[Shard(0), Shard(0)],
]
TWO_D_TO_TWO_D_PLACEMENTS = []
for p1 in TWO_D_PLACEMENTS:
for p2 in TWO_D_PLACEMENTS:
if p1 != p2:
TWO_D_TO_TWO_D_PLACEMENTS.append((p1, p2))
class TestDTensorReshardPlacementChange(DTensorTestBase):
"""
Test DCP reshard for DTensor with placements changes and without world_size change and mesh_tensor change.
"""
@with_comms
@skip_if_lt_x_gpu(2)
@with_temp_dir
def test_1d_to_1d_reshard_placement_change(self) -> None:
CHECKPOINT_DIR = self.temp_dir
for one_d_to_one_d_placements in ONE_D_TO_ONE_D_PLACEMENTS:
original_placement, new_placement = one_d_to_one_d_placements
global_tensor = torch.arange(16, dtype=torch.float).view(4, 4)
mesh_shape = (self.world_size,)
device_mesh = init_device_mesh(self.device_type, mesh_shape)
dtensor = distribute_tensor(
global_tensor, device_mesh, placements=original_placement
)
state_dict_to_save = {"dtensor": dtensor}
dist_cp.save(
state_dict=state_dict_to_save,
storage_writer=dist_cp.FileSystemWriter(path=CHECKPOINT_DIR),
planner=dist_cp.DefaultSavePlanner(),
)
zero_dtensor = zeros(
[4, 4], device_mesh=device_mesh, placements=new_placement
)
state_dict_to_load = {"dtensor": zero_dtensor}
dist_cp.load(
state_dict=state_dict_to_load,
storage_reader=dist_cp.FileSystemReader(CHECKPOINT_DIR),
planner=dist_cp.DefaultLoadPlanner(),
)
# materialzie the whole tensor to compare with the original global_tensor
state_dict_to_load["dtensor"] = state_dict_to_load["dtensor"].redistribute(
device_mesh,
placements=[Replicate()],
)
self.assertEqual(global_tensor, state_dict_to_load["dtensor"].to_local())
# redistribute the tensor back to its original placement for comparison.
state_dict_to_load["dtensor"] = state_dict_to_load["dtensor"].redistribute(
device_mesh,
placements=original_placement,
)
self.assertEqual(
state_dict_to_save["dtensor"].to_local(),
state_dict_to_load["dtensor"].to_local(),
)
@with_comms
@skip_if_lt_x_gpu(4)
@with_temp_dir
def test_2d_to_2d_reshard_placement_change(self) -> None:
CHECKPOINT_DIR = self.temp_dir
for two_d_to_two_d_placements in TWO_D_TO_TWO_D_PLACEMENTS:
original_placement, new_placement = two_d_to_two_d_placements
global_tensor = torch.arange(16, dtype=torch.float).view(4, 4)
mesh_shape = (2, self.world_size // 2)
mesh_2d = init_device_mesh(self.device_type, mesh_shape)
dtensor = distribute_tensor(
global_tensor,
mesh_2d,
placements=original_placement,
)
state_dict_to_save = {"dtensor": dtensor}
dist_cp.save(
state_dict=state_dict_to_save,
storage_writer=dist_cp.FileSystemWriter(path=CHECKPOINT_DIR),
planner=dist_cp.DefaultSavePlanner(),
)
zero_dtensor = zeros([4, 4], device_mesh=mesh_2d, placements=new_placement)
state_dict_to_load = {"dtensor": zero_dtensor}
dist_cp.load(
state_dict=state_dict_to_load,
storage_reader=dist_cp.FileSystemReader(CHECKPOINT_DIR),
planner=dist_cp.DefaultLoadPlanner(),
)
state_dict_to_load["dtensor"] = state_dict_to_load["dtensor"].redistribute(
mesh_2d,
placements=[Replicate(), Replicate()],
)
self.assertEqual(global_tensor, state_dict_to_load["dtensor"].to_local())
state_dict_to_load["dtensor"] = state_dict_to_load["dtensor"].redistribute(
mesh_2d,
placements=original_placement,
)
self.assertEqual(
state_dict_to_save["dtensor"].to_local(),
state_dict_to_load["dtensor"].to_local(),
)
class TestDTensorReshardMeshChange(DTensorTestBase):
"""
Test DCP reshard for DTensor with placements changes and mesh_tensor change.
"""
@with_comms
@with_temp_dir
@skip_if_lt_x_gpu(2)
def test_1d_to_2d_reshard_mesh_change(self) -> None:
CHECKPOINT_DIR = self.temp_dir
for placements_1d in ONE_D_PLACEMENTS:
global_tensor = torch.arange(16, dtype=torch.float).view(4, 4)
mesh_shape = (self.world_size,)
mesh_1d = init_device_mesh(self.device_type, mesh_shape)
dtensor = distribute_tensor(
global_tensor, mesh_1d, placements=placements_1d
)
state_dict_to_save = {"dtensor": dtensor}
dist_cp.save(
state_dict=state_dict_to_save,
storage_writer=dist_cp.FileSystemWriter(path=CHECKPOINT_DIR),
planner=dist_cp.DefaultSavePlanner(),
)
for placements_2d in TWO_D_PLACEMENTS:
mesh_shape = (2, self.world_size // 2)
mesh_2d = init_device_mesh(self.device_type, mesh_shape)
zero_dtensor = zeros(
[4, 4], device_mesh=mesh_2d, placements=placements_2d
)
state_dict_to_load = {"dtensor": zero_dtensor}
dist_cp.load(
state_dict=state_dict_to_load,
storage_reader=dist_cp.FileSystemReader(CHECKPOINT_DIR),
planner=dist_cp.DefaultLoadPlanner(),
)
# materialzie the whole tensor to compare with the original global_tensor
state_dict_to_load["dtensor"] = state_dict_to_load[
"dtensor"
].redistribute(
mesh_2d,
placements=[Replicate(), Replicate()],
)
self.assertEqual(
global_tensor, state_dict_to_load["dtensor"].to_local()
)
@with_comms
@with_temp_dir
@skip_if_lt_x_gpu(4)
def test_2d_to_1d_reshard_mesh_change(self) -> None:
CHECKPOINT_DIR = self.temp_dir
for placements_2d in TWO_D_PLACEMENTS:
global_tensor = torch.arange(16, dtype=torch.float).view(4, 4)
mesh_shape = (2, self.world_size // 2)
mesh_2d = init_device_mesh(self.device_type, mesh_shape)
dtensor = distribute_tensor(
global_tensor, mesh_2d, placements=placements_2d
)
state_dict_to_save = {"dtensor": dtensor}
dist_cp.save(
state_dict=state_dict_to_save,
storage_writer=dist_cp.FileSystemWriter(path=CHECKPOINT_DIR),
planner=dist_cp.DefaultSavePlanner(),
)
for placements_1d in ONE_D_PLACEMENTS:
mesh_shape = (self.world_size,)
mesh_1d = init_device_mesh(self.device_type, mesh_shape)
zero_dtensor = zeros(
[4, 4], device_mesh=mesh_1d, placements=placements_1d
)
state_dict_to_load = {"dtensor": zero_dtensor}
dist_cp.load(
state_dict=state_dict_to_load,
storage_reader=dist_cp.FileSystemReader(CHECKPOINT_DIR),
planner=dist_cp.DefaultLoadPlanner(),
)
# materialzie the whole tensor to compare with the original global_tensor
state_dict_to_load["dtensor"] = state_dict_to_load[
"dtensor"
].redistribute(
mesh_1d,
placements=[Replicate()],
)
self.assertEqual(
global_tensor, state_dict_to_load["dtensor"].to_local()
)
@with_comms
@with_temp_dir
@skip_if_lt_x_gpu(2)
def test_dtensor_checkpoint_resharding_with_empty_shard(self):
"""
Test dtensor checkpoint resharding with dtensor containing empty shards.
"""
tensor = torch.rand(1).cuda()
mesh = init_device_mesh(self.device_type, (self.world_size,))
dtensor = distribute_tensor(tensor, mesh, [Shard(0)])
ref_state_dict = {"dtensor": dtensor}
dist_cp.save(
state_dict=ref_state_dict,
storage_writer=dist_cp.FileSystemWriter(path=self.temp_dir),
)
tensor = torch.rand(1).cuda()
mesh_2 = init_device_mesh(self.device_type, (2, self.world_size // 2))
dtensor = distribute_tensor(tensor, mesh_2, [Shard(0), Shard(0)])
state_dict = {"dtensor": dtensor}
dist_cp.load(
state_dict=state_dict,
storage_reader=dist_cp.FileSystemReader(self.temp_dir),
)
# TODO: Add a assertEqual for ref_state_dict["dtensor"].full_tensor()
# and state_dict["dtensor"].full_tensor() after we fix the size mismatch
# issue for un-even sharding dtensor.
# TODO: Add dtensor resharding test when world size changes.
if __name__ == "__main__":
run_tests()
|
