test_engine.py 2.52 KB
Newer Older
1
2
3
4
5
6
7
8
9
from functools import partial

import colossalai
import pytest
import torch.multiprocessing as mp
from colossalai.amp import AMP_TYPE
from colossalai.core import global_context as gpc
from colossalai.utils import free_port
from tests.components_to_test.registry import non_distributed_component_funcs
10
from colossalai.testing import parameterize, rerun_on_exception
11
12
13
14
15
16

CONFIG = dict(parallel=dict(pipeline=dict(size=1), tensor=dict(size=1, mode=None)),
              fp16=dict(mode=None),
              clip_grad_norm=1.0)


17
18
19
@parameterize('model_name', ['repeated_computed_layers', 'resnet18', 'repeated_computed_layers'])
@parameterize('amp_mode', [AMP_TYPE.APEX, AMP_TYPE.TORCH, AMP_TYPE.NAIVE, None])
def run_train(model_name, amp_mode):
ver217's avatar
ver217 committed
20
    # FIXME: test bert
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
    get_components_func = non_distributed_component_funcs.get_callable(model_name)
    gpc.config.fp16['mode'] = amp_mode
    model_builder, train_dataloader, _, optimizer_class, criterion = get_components_func()

    model = model_builder(checkpoint=False)
    engine, train_dataloader, *args = colossalai.initialize(model=model,
                                                            optimizer=optimizer_class(model.parameters(), lr=1e-3),
                                                            criterion=criterion,
                                                            train_dataloader=train_dataloader)

    try:
        engine.train()
        for data, label in train_dataloader:
            engine.zero_grad()
            data = data.cuda()
            label = label.cuda()
            if criterion:
                output = engine(data)
                loss = engine.criterion(output, label)
            else:
                loss = engine(data, label)
            engine.backward(loss)
            engine.step()
            break
    except IndexError:
        # if using apex amp, NetWithRepeatedlyComputedLayers will raise an index out of range issue
        # the following check fails in apex
        # if cached_x.grad_fn.next_functions[1][0].variable is not x:
        pass
50
51
52
53


def run_engine(rank, world_size, port):
    # init dist env
54
55
    colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
    run_train()
56
57
58


@pytest.mark.dist
59
@rerun_on_exception(exception_type=mp.ProcessRaisedException, pattern=".*Address already in use.*")
60
def test_engine():
61
    world_size = 2
62
63
64
65
66
67
    run_func = partial(run_engine, world_size=world_size, port=free_port())
    mp.spawn(run_func, nprocs=world_size)


if __name__ == '__main__':
    test_engine()