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
|
# Owner(s): ["oncall: distributed"]
import sys
import torch
from torch.distributed._shard.checkpoint.planner import LoadItemType, WriteItemType
from torch.distributed._shard.sharded_tensor import (
Shard,
ShardMetadata,
ShardedTensor,
ShardedTensorMetadata,
)
from torch.distributed._shard.sharded_tensor.metadata import TensorProperties
from torch.testing._internal.common_utils import (
TestCase,
TEST_WITH_DEV_DBG_ASAN,
run_tests,
)
from torch.distributed._shard.checkpoint.metadata import BytesStorageMetadata, MetadataIndex, TensorStorageMetadata
from torch.testing._internal.distributed.distributed_utils import (
with_fake_comms,
with_dist
)
from torch.distributed._shard.checkpoint.default_planner import (
create_default_global_save_plan,
create_default_local_save_plan,
create_default_local_load_plan,
_create_default_local_metadata
)
if TEST_WITH_DEV_DBG_ASAN:
print(
"Skip dev-asan as torch + multiprocessing spawn have known issues",
file=sys.stderr,
)
sys.exit(0)
def create_sharded_tensor(rank, world_size, shards_per_rank, shard_size=8):
shards_metadata = []
local_shards = []
for idx in range(0, world_size * shards_per_rank):
shard_rank = idx // shards_per_rank
shard_md = ShardMetadata(shard_offsets=[idx * shard_size], shard_sizes=[shard_size], placement=f"rank:{shard_rank}/cpu")
shards_metadata.append(shard_md)
if shard_rank == rank:
shard = Shard.from_tensor_and_offsets(
torch.rand(*shard_md.shard_sizes),
shard_offsets=shard_md.shard_offsets,
rank=rank
)
local_shards.append(shard)
sharded_tensor_md = ShardedTensorMetadata(
shards_metadata=shards_metadata,
size=torch.Size([shard_size * len(shards_metadata)]),
tensor_properties=TensorProperties.create_from_tensor(torch.zeros(1))
)
return ShardedTensor._init_from_local_shards_and_global_metadata(
local_shards=local_shards,
sharded_tensor_metadata=sharded_tensor_md
)
class TestSavePlan(TestCase):
@with_fake_comms(rank=1, world_size=4)
def test_local_plan(self):
tensor = torch.rand(10)
val = [1, 2, 3]
st = create_sharded_tensor(rank=1, world_size=4, shards_per_rank=1)
state_dict = {
"tensor": tensor,
"value": val,
"st": st
}
plan = create_default_local_save_plan(state_dict, False)
self.assertEqual(1, len(plan.items))
wi = plan.items[0]
self.assertEqual(wi.index, MetadataIndex("st", [8]))
self.assertEqual(wi.type, WriteItemType.SHARD)
self.assertEqual(wi.tensor_data.size, st.size())
self.assertEqual(wi.tensor_data.properties, TensorProperties.create_from_tensor(torch.zeros(1)))
self.assertEqual(wi.tensor_data.chunk.offsets, torch.Size([8]))
self.assertEqual(wi.tensor_data.chunk.sizes, torch.Size([8]))
# Coordinator rank, should include replicated items as well
plan = create_default_local_save_plan(state_dict, True)
self.assertEqual(3, len(plan.items))
tensor_wi = next(wi for wi in plan.items if wi.type == WriteItemType.TENSOR)
self.assertEqual(tensor_wi.index, MetadataIndex("tensor", [0]))
self.assertEqual(tensor_wi.tensor_data.size, tensor.size())
self.assertEqual(tensor_wi.tensor_data.properties, TensorProperties.create_from_tensor(tensor))
self.assertEqual(tensor_wi.tensor_data.chunk.offsets, torch.Size([0]))
self.assertEqual(tensor_wi.tensor_data.chunk.sizes, torch.Size([10]))
bytes_wi = next(wi for wi in plan.items if wi.type == WriteItemType.BYTE_IO)
self.assertEqual(bytes_wi.index, MetadataIndex("value"))
self.assertIsNone(bytes_wi.tensor_data)
def test_global_plan(self):
def create_data(rank):
with with_dist(rank=rank, world_size=4):
tensor = torch.rand(10)
val = [1, 2, 3]
st = create_sharded_tensor(rank=rank, world_size=4, shards_per_rank=1)
state_dict = {
"tensor": tensor,
"value": val,
"st": st
}
return create_default_local_save_plan(state_dict, rank == 0)
all_plans = [create_data(0), create_data(1), create_data(2), create_data(3)]
final_plans, metadata = create_default_global_save_plan(all_plans=all_plans)
# The default global plan updates all indexes to include hints
for new_plan, old_plan in zip(final_plans, all_plans):
for new_item, old_item in zip(new_plan.items, old_plan.items):
self.assertEqual(new_item.index, old_item.index)
self.assertEqual(new_item.type, old_item.type)
self.assertEqual(new_item.tensor_data, old_item.tensor_data)
self.assertIn(new_item.index.fqn, metadata.state_dict_metadata)
item_md = metadata.state_dict_metadata[new_item.index.fqn]
if new_item.type == WriteItemType.BYTE_IO:
self.assertTrue(isinstance(item_md, BytesStorageMetadata))
else:
self.assertTrue(isinstance(item_md, TensorStorageMetadata))
self.assertEqual(item_md.size, old_item.tensor_data.size)
self.assertEqual(item_md.properties, old_item.tensor_data.properties)
self.assertIsNotNone(new_item.index.index)
# Make sure the hint is correct
self.assertEqual(item_md.chunks[new_item.index.index], old_item.tensor_data.chunk)
def test_local_load_plan(self):
def create_state_dict(rank):
with with_dist(rank=rank, world_size=4):
tensor = torch.rand(10)
val = [1, 2, 3]
st = create_sharded_tensor(rank=rank, world_size=4, shards_per_rank=1)
return {
"tensor": tensor,
"value": val,
"st": st
}
state_dict = create_state_dict(1)
metadata = _create_default_local_metadata(state_dict)
load_plan = create_default_local_load_plan(state_dict, metadata)
# This will create 3 entries
self.assertEqual(3, len(load_plan.items))
st_item = next(ri for ri in load_plan.items if ri.dest_index.fqn == "st")
tensor_item = next(ri for ri in load_plan.items if ri.dest_index.fqn == "tensor")
bytes_item = next(ri for ri in load_plan.items if ri.dest_index.fqn == "value")
self.assertEqual(st_item.type, LoadItemType.TENSOR)
# This is an exact copy
self.assertEqual(st_item.dest_index, MetadataIndex("st", [8]))
self.assertEqual(st_item.dest_offsets, torch.Size([0]))
self.assertEqual(st_item.storage_index, MetadataIndex("st", [8]))
self.assertEqual(st_item.storage_offsets, torch.Size([0]))
self.assertEqual(st_item.lengths, torch.Size([8]))
self.assertEqual(tensor_item.type, LoadItemType.TENSOR)
self.assertEqual(tensor_item.dest_index, MetadataIndex("tensor", [0]))
self.assertEqual(tensor_item.dest_offsets, torch.Size([0]))
self.assertEqual(tensor_item.storage_index, MetadataIndex("tensor", [0]))
self.assertEqual(tensor_item.storage_offsets, torch.Size([0]))
self.assertEqual(tensor_item.lengths, torch.Size([10]))
self.assertEqual(bytes_item.type, LoadItemType.BYTE_IO)
self.assertEqual(bytes_item.dest_index, MetadataIndex("value"))
def test_load_with_resharding(self):
def create_state_dict(rank, world_size):
with with_dist(rank=rank, world_size=world_size):
return {
"st": create_sharded_tensor(
rank=rank,
world_size=world_size,
shards_per_rank=1,
shard_size=128 // world_size,
)
}
# Rank 1 has a 16 bytes shard from [16, 32[
world8_state_dict = create_state_dict(rank=1, world_size=8)
world8_metadata = _create_default_local_metadata(world8_state_dict)
# Rank 1 has a 32 bytes shard from [32, 64[
world4_state_dict = create_state_dict(rank=1, world_size=4)
world4_metadata = _create_default_local_metadata(world4_state_dict)
# First scenario, going from world=8 to world=4, need to load 2 shards
# Each 4-world shard has 32 elements, so it needs to load 2 shards
load_plan = create_default_local_load_plan(world4_state_dict, world8_metadata)
self.assertEqual(2, len(load_plan.items))
low_ri = next(ri for ri in load_plan.items if ri.dest_offsets == torch.Size([0]))
high_ri = next(ri for ri in load_plan.items if ri.dest_offsets == torch.Size([16]))
self.assertEqual(low_ri.storage_index, MetadataIndex("st", [32]))
self.assertEqual(low_ri.storage_offsets, torch.Size([0]))
self.assertEqual(low_ri.dest_index, MetadataIndex("st", [32]))
self.assertEqual(low_ri.dest_offsets, torch.Size([0]))
self.assertEqual(low_ri.lengths, torch.Size([16]))
self.assertEqual(high_ri.storage_index, MetadataIndex("st", [48]))
self.assertEqual(high_ri.storage_offsets, torch.Size([0]))
self.assertEqual(high_ri.dest_index, MetadataIndex("st", [32]))
self.assertEqual(high_ri.dest_offsets, torch.Size([16]))
self.assertEqual(high_ri.lengths, torch.Size([16]))
# Second scenario, going from world=4 to world=8, need to load half of 1 shard
# rank1 on 8-world needs to load the upper half of the rank0 4-world shard
load_plan = create_default_local_load_plan(world8_state_dict, world4_metadata)
self.assertEqual(1, len(load_plan.items))
ri = load_plan.items[0]
self.assertEqual(ri.storage_index, MetadataIndex("st", [0]))
self.assertEqual(ri.storage_offsets, torch.Size([16]))
self.assertEqual(ri.dest_index, MetadataIndex("st", [16]))
self.assertEqual(ri.dest_offsets, torch.Size([0]))
self.assertEqual(ri.lengths, torch.Size([16]))
def test_load_with_world_size_diff_by_one(self):
def create_state_dict(rank, world_size):
with with_dist(rank=rank, world_size=world_size):
return {
"st": create_sharded_tensor(
rank=rank,
world_size=world_size,
shards_per_rank=1,
shard_size=120 // world_size,
)
}
# rank 1 has a 30 bytes shard from [30, 60[
world4_state_dict = create_state_dict(rank=1, world_size=4)
world4_metadata = _create_default_local_metadata(world4_state_dict)
# rank 1 has a 40 bytes shard from [40, 80[
world3_state_dict = create_state_dict(rank=1, world_size=3)
load_plan = create_default_local_load_plan(world3_state_dict, world4_metadata)
self.assertEqual(2, len(load_plan.items))
# this is [30, 60] to load [40, 60]
low_ri = next(ri for ri in load_plan.items if ri.dest_offsets == torch.Size([0]))
# this is [60, 90] to load [60, 80]
high_ri = next(ri for ri in load_plan.items if ri.dest_offsets == torch.Size([20]))
self.assertEqual(low_ri.storage_index, MetadataIndex("st", [30]))
self.assertEqual(low_ri.storage_offsets, torch.Size([10]))
self.assertEqual(low_ri.dest_index, MetadataIndex("st", [40]))
self.assertEqual(low_ri.dest_offsets, torch.Size([0]))
self.assertEqual(low_ri.lengths, torch.Size([20]))
self.assertEqual(high_ri.storage_index, MetadataIndex("st", [60]))
self.assertEqual(high_ri.storage_offsets, torch.Size([0]))
self.assertEqual(high_ri.dest_index, MetadataIndex("st", [40]))
self.assertEqual(high_ri.dest_offsets, torch.Size([20]))
self.assertEqual(high_ri.lengths, torch.Size([20]))
if __name__ == "__main__":
run_tests()
|
