test_3d.py 2.04 KB
Newer Older
zbian's avatar
zbian committed
1
2
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
アマデウス's avatar
アマデウス committed
3
4
from functools import partial

Frank Lee's avatar
Frank Lee committed
5
6
7
import pytest
import torch
import torch.multiprocessing as mp
アマデウス's avatar
アマデウス committed
8
9
from colossalai.core import global_context as gpc
from colossalai.initialize import launch
10
from colossalai.logging import disable_existing_loggers
11
from colossalai.utils import free_port
12
from colossalai.testing import rerun_if_address_is_in_use, skip_if_not_enough_gpus
13
14
15
16
17
from checks_3d.check_layer_3d import (check_classifier_given_embed_weight, check_classifier_no_given_weight,
                                      check_embed, check_layernorm, check_linear, check_loss, check_patch_embed,
                                      check_vocab_parallel_classifier_given_embed_weight,
                                      check_vocab_parallel_classifier_no_given_weight, check_vocab_parallel_embed,
                                      check_vocab_parallel_loss)
zbian's avatar
zbian committed
18

アマデウス's avatar
アマデウス committed
19
20
21
22
23
24
25
CONFIG = dict(
    parallel=dict(
        pipeline=1,
        tensor=dict(mode='3d', size=8),
    ),
    seed=42,
)
zbian's avatar
zbian committed
26
27
28


def check_layer():
アマデウス's avatar
アマデウス committed
29
30
    check_linear()
    check_layernorm()
31
32
33
34
35
36
37
38
39
    check_classifier_no_given_weight()
    check_vocab_parallel_classifier_no_given_weight()
    check_classifier_given_embed_weight()
    check_vocab_parallel_classifier_given_embed_weight()
    check_embed()
    check_patch_embed()
    check_vocab_parallel_embed()
    check_loss()
    check_vocab_parallel_loss()
zbian's avatar
zbian committed
40

41

42
def check_layer_and_operation(rank, world_size, port):
43
    disable_existing_loggers()
44
    launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
45
46
47
    torch.backends.cuda.matmul.allow_tf32 = False
    torch.backends.cudnn.allow_tf32 = False
    torch.backends.cudnn.deterministic = True
zbian's avatar
zbian committed
48
    check_layer()
Frank Lee's avatar
Frank Lee committed
49
50
51
52
53
    gpc.destroy()
    torch.cuda.empty_cache()


@pytest.mark.dist
54
@skip_if_not_enough_gpus(min_gpus=8)
55
@rerun_if_address_is_in_use()
Frank Lee's avatar
Frank Lee committed
56
57
def test_3d():
    world_size = 8
58
    run_func = partial(check_layer_and_operation, world_size=world_size, port=free_port())
Frank Lee's avatar
Frank Lee committed
59
    mp.spawn(run_func, nprocs=world_size)
zbian's avatar
zbian committed
60
61
62


if __name__ == '__main__':
Frank Lee's avatar
Frank Lee committed
63
    test_3d()