File size: 6,891 Bytes
23bd7af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from commons import print_separator
from commons import initialize_distributed
import mpu
import torch
import sys
sys.path.append("../..")


def test_set_cuda_rng_state(tensor_model_parallel_size):

    if torch.distributed.get_rank() == 0:
        print('> testing set_rng_state with size {} ...'.
              format(tensor_model_parallel_size))

    mpu.initialize_model_parallel(tensor_model_parallel_size)
    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()

    size = 123
    seed = 1234
    torch.cuda.manual_seed(1234)
    tensor = torch.cuda.FloatTensor(size)

    # Get the state
    rng_state = torch.cuda.get_rng_state()
    rng_state_copy = rng_state.clone()

    # Do some stuff.
    for _ in range(5):
        torch.randn(size, out=tensor)
    result_1 = tensor.clone()

    assert rng_state.sub(rng_state_copy).max() == 0
    assert torch.cuda.get_rng_state().sub(rng_state_copy).max() > 0

    # State should be different.
    new_rng_state = torch.cuda.get_rng_state()
    max_diff = new_rng_state.sub(rng_state).max()
    print('   max diff in rng state (should be non-zero) on global rank {}: {}'.
          format(torch.distributed.get_rank(), max_diff))
    assert max_diff > 0

    # Reset the rng state and do the same stuff.
    mpu.random._set_cuda_rng_state(rng_state)
    for _ in range(5):
        torch.randn(size, out=tensor)
    mpu.random._set_cuda_rng_state(rng_state)
    for _ in range(5):
        torch.randn(size, out=tensor)
    result_2 = tensor.clone()

    # Results should be the same
    error = result_2.sub(result_1).abs().max()
    print('   max error in generated tensors (should be zero) on '
          'global rank {}: {}'.format(torch.distributed.get_rank(), error))
    assert error < 1.0e-6

    # Input state should have remained intact.
    error = rng_state.sub(rng_state_copy).max()
    print('   max error in rng state (should be zero) on global rank {}: {}'.
          format(torch.distributed.get_rank(), error))
    assert error == 0

    # Reset groups
    mpu.destroy_model_parallel()

    torch.distributed.barrier()
    if torch.distributed.get_rank() == 0:
        print('>> passed the test :-)')


def test_cuda_rng_tracker(tensor_model_parallel_size):

    if torch.distributed.get_rank() == 0:
        print('> testing cuda rng tracker with size {} ...'.
              format(tensor_model_parallel_size))

    mpu.initialize_model_parallel(tensor_model_parallel_size)
    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()

    seed_1 = 1234
    seed_2 = 4321
    size = [12, 21]
    tensor = torch.cuda.FloatTensor(size)

    # Set to seed_1 and generate two tensors.
    torch.cuda.manual_seed(seed_1)
    torch.randn(size, out=tensor)
    target_11 = tensor.clone()
    torch.randn(size, out=tensor)
    target_12 = tensor.clone()

    # Set to seed_2 and generate two tensors.
    torch.cuda.manual_seed(seed_2)
    torch.randn(size, out=tensor)
    target_21 = tensor.clone()
    torch.randn(size, out=tensor)
    target_22 = tensor.clone()

    # Now if we interleave seed_1 and seed_2,
    # we should still get the same tensors
    torch.cuda.manual_seed(seed_1)
    mpu.get_cuda_rng_tracker().add('test', seed_2)

    torch.randn(size, out=tensor)
    result_11 = tensor.clone()

    with mpu.get_cuda_rng_tracker().fork('test'):
        torch.randn(size, out=tensor)
        result_21 = tensor.clone()

    torch.randn(size, out=tensor)
    result_12 = tensor.clone()

    with mpu.get_cuda_rng_tracker().fork('test'):
        torch.randn(size, out=tensor)
        result_22 = tensor.clone()

    diff = result_11.sub(result_21).abs().max()
    diff = min(diff, result_12.sub(result_22).abs().max())
    print('   max diff in generated tensors (should be non-zero) on '
          'global rank {}: {}'.format(torch.distributed.get_rank(), diff))
    assert diff > 1.0e-6
    error = max(result_11.sub(target_11).abs().max(),
                result_12.sub(target_12).abs().max())
    error = max(error, result_21.sub(target_21).abs().max())
    error = max(error, result_22.sub(target_22).abs().max())
    print('   max error in generated tensors (should be zero) on '
          'global rank {}: {}'.format(torch.distributed.get_rank(), error))
    assert error < 1.0e-6

    # Reset the tracker
    mpu.get_cuda_rng_tracker().reset()

    # Reset groups
    mpu.destroy_model_parallel()

    torch.distributed.barrier()
    if torch.distributed.get_rank() == 0:
        print('>> passed the test :-)')


def test_model_parallel_cuda_manual_seed(tensor_model_parallel_size):

    if torch.distributed.get_rank() == 0:
        print('> testing model parallel cuda manual seed with size {} ...'.
              format(tensor_model_parallel_size))

    mpu.initialize_model_parallel(tensor_model_parallel_size)
    tensor_model_parallel_size = mpu.get_tensor_model_parallel_world_size()

    mpu.model_parallel_cuda_manual_seed(12345)
    assert torch.cuda.initial_seed() == 12345
    with mpu.get_cuda_rng_tracker().fork():
        assert torch.cuda.initial_seed() == (12345 + 2718 +
                                             mpu.get_tensor_model_parallel_rank())

    # Reset the tracker
    mpu.get_cuda_rng_tracker().reset()

    # Reset groups
    mpu.destroy_model_parallel()

    torch.distributed.barrier()
    if torch.distributed.get_rank() == 0:
        print('>> passed the test :-)')


if __name__ == '__main__':

    initialize_distributed()
    world_size = torch.distributed.get_world_size()

    tensor_model_parallel_size = 1
    while tensor_model_parallel_size <= world_size:
        print_separator('test set rng state')
        test_set_cuda_rng_state(tensor_model_parallel_size)
        tensor_model_parallel_size *= 2

    tensor_model_parallel_size = 1
    while tensor_model_parallel_size <= world_size:
        print_separator('test cuda rng tracker')
        test_cuda_rng_tracker(tensor_model_parallel_size)
        tensor_model_parallel_size *= 2

    tensor_model_parallel_size = 1
    while tensor_model_parallel_size <= world_size:
        print_separator('test model parallel cuda manual seed')
        test_model_parallel_cuda_manual_seed(tensor_model_parallel_size)
        tensor_model_parallel_size *= 2