test_ddp_state_dict.py 2.1 KB
Newer Older
1
from collections import OrderedDict
2

3
4
import pytest
import torch
5
6
7
8

import colossalai
from colossalai.nn.parallel import ColoDDP
from colossalai.tensor import ColoParameter, ProcessGroup
9
from colossalai.testing import rerun_if_address_is_in_use, spawn
10
11
from colossalai.utils.cuda import get_current_device
from colossalai.zero import ColoInitContext
12
13
14
15
16
17
from tests.components_to_test.registry import non_distributed_component_funcs


def check_state_dict_equal(state_dict: OrderedDict, other_state_dict: OrderedDict):
    for (k1, t1), (k2, t2) in zip(state_dict.items(), other_state_dict.items()):
        assert k1 == k2
18
19
20
21
22
23

        if t1.device != t2.device:
            temp_t2 = t2.to(t1.device)
        else:
            temp_t2 = t2

24
25
26
        assert torch.equal(t1, temp_t2), "\t{}\n\t{}".format(t1, temp_t2)


27
def init_ddp(module: torch.nn.Module) -> ColoDDP:
28
29
    pg = ProcessGroup()
    return ColoDDP(module, process_group=pg)
30
31


32
33
def run_ddp_state_dict():
    get_components_func = non_distributed_component_funcs.get_callable('gpt2')
34
35
36
37
    model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
    torch_model = model_builder().cuda()
    with ColoInitContext(device=get_current_device()):
        model = model_builder()
38
    model = init_ddp(model)
39
    torch_state_dict = torch_model.state_dict()
40

41
42
43
    for param in model.parameters():
        if isinstance(param, ColoParameter):
            assert param.get_process_group() is not None
44
    model.load_state_dict(torch_state_dict)
45
46
47
48
49

    for param in model.parameters():
        if isinstance(param, ColoParameter):
            assert param.get_process_group() is not None

50
51
52
53
54
55
    state_dict = model.state_dict()
    check_state_dict_equal(torch_state_dict, state_dict)


def run_dist(rank, world_size, port):
    colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
56
    run_ddp_state_dict()
57
58
59
60
61
62


@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 2])
@rerun_if_address_is_in_use()
def test_state_dict(world_size):
63
    spawn(run_dist, world_size)
64
65
66
67


if __name__ == '__main__':
    test_state_dict(2)