colossalai

1
import pytest
2
import torch
3
import torch.distributed as dist
4
from torch.nn.parallel import DistributedDataParallel as DDP
5
from torch.testing import assert_close
6

7
import colossalai
8
from colossalai.legacy.amp import convert_to_apex_amp
9
from colossalai.nn.optimizer import HybridAdam
10
from colossalai.testing import DummyDataloader, parameterize, rerun_if_address_is_in_use, spawn
11
from colossalai.utils import set_seed
12
from colossalai.zero import GeminiDDP, GeminiOptimizer
13
from colossalai.zero.gemini.chunk import search_chunk_configuration
14
from tests.kit.model_zoo import model_zoo, run_fwd_bwd
15

16
PLACEMENT_CONFIGS = [
17
    {
18
        "placement_policy": "static",
19
        "shard_param_frac": 0.0,
20
        "offload_optim_frac": 0.0,
21
        "offload_param_frac": 0.0,
22
    },  # zero2
23
    {
24
        "placement_policy": "static",
25
        "shard_param_frac": 0.0,
26
        "offload_optim_frac": 1.0,
27
        "offload_param_frac": 0.0,
28
    },  # zero2-offload
29
    {
30
        "placement_policy": "static",
31
        "shard_param_frac": 0.0,
32
        "offload_optim_frac": 0.5,
33
        "offload_param_frac": 0.0,
34
    },  # zero2-offload-half
35
    {"placement_policy": "auto"},
36
]
37

38

39
def check_param(model: GeminiDDP, torch_model: torch.nn.Module):
40
    zero_dict = model.state_dict(only_rank_0=False)
41
    torch_dict = torch_model.state_dict()
42

43
    for key, value in torch_dict.items():
44
        # key is 'module.model.PARAMETER', so we truncate it
45
        key = key[7:]
46
        assert key in zero_dict, "{} not in ZeRO dictionary.".format(key)
47
        temp_zero_value = zero_dict[key].to(device=value.device, dtype=value.dtype)
48
        # debug_print([0], "max range: ", key, torch.max(torch.abs(value - temp_zero_value)))
49
        assert_close(value, temp_zero_value, rtol=1e-3, atol=4e-3)
50

51

52
@parameterize("placement_config", PLACEMENT_CONFIGS)
53
@parameterize("model_name", ["transformers_gpt_lm"])
54
@parameterize("master_weights", [True, False])
55
def exam_grad_clipping(placement_config, model_name: str, master_weights: bool):
56
    set_seed(1912)
57
    model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
58
        iter(model_zoo.get_sub_registry(model_name).values())
59
    )
60

61
    torch_model = model_builder().cuda()
62
    amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=32)
63
    torch_optim = torch.optim.Adam(torch_model.parameters(), lr=1e-3)
64
    torch_model, torch_optim = convert_to_apex_amp(torch_model, torch_optim, amp_config)
65
    torch_model = DDP(torch_model, device_ids=[dist.get_rank()])
66

67
    model = model_builder()
68

69
    for torch_p, p in zip(torch_model.parameters(), model.parameters()):
70
        p.data.copy_(torch_p.data)
71

72
    world_size = torch.distributed.get_world_size()
73
    config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
74
    config_dict[world_size]["chunk_size"] = 5000
75
    config_dict[world_size]["keep_gathered"] = False
76
    if placement_config["placement_policy"] != "cuda":
77
        init_device = torch.device("cpu")
78
    else:
79
        init_device = None
80

81
    model = GeminiDDP(
82
        model,
83
        chunk_config_dict=config_dict,
84
        chunk_init_device=init_device,
85
        pin_memory=True,
86
        master_weights=master_weights,
87
        **placement_config,
88
    )
89

90
    optimizer = HybridAdam(model.parameters(), lr=1e-3)
91
    zero_optim = GeminiOptimizer(optimizer, model, initial_scale=32, max_norm=1.0)
92

93
    model.train()
94
    torch_model.train()
95

96
    set_seed(dist.get_rank() * 3 + 128)
97
    train_dataloader = DummyDataloader(data_gen_fn)
98
    for i, data in enumerate(train_dataloader):
99
        if i > 2:
100
            break
101
        data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
102

103
        zero_optim.zero_grad()
104
        torch_optim.zero_grad()
105

106
        run_fwd_bwd(torch_model, data, output_transform_fn, loss_fn, optimizer=torch_optim)
107
        run_fwd_bwd(model, data, output_transform_fn, loss_fn, optimizer=zero_optim)
108

109
        import apex.amp as apex_amp
110

111
        torch.nn.utils.clip_grad_norm_(apex_amp.master_params(torch_optim), 1.0)
112
        torch_optim.step()
113
        zero_optim.step()
114

115
        if master_weights:
116
            check_param(model, torch_model)
117

118

119
def run_dist(rank, world_size, port):
120
    config = {}
121
    colossalai.launch(config=config, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
122
    exam_grad_clipping()
123

124

125
@pytest.mark.dist
126
@pytest.mark.parametrize("world_size", [1, 2])
127
@rerun_if_address_is_in_use()
128
def test_grad_clip(world_size):
129
    spawn(run_dist, world_size)
130

131

132
if __name__ == "__main__":
133
    test_grad_clip(2)
134