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.accelerator import get_accelerator
9
from colossalai.legacy.amp import convert_to_apex_amp
10
from colossalai.nn.optimizer import HybridAdam
11
from colossalai.testing import DummyDataloader, parameterize, rerun_if_address_is_in_use, spawn
12
from colossalai.utils import set_seed
13
from colossalai.zero import GeminiDDP, GeminiOptimizer
14
from colossalai.zero.gemini.chunk import search_chunk_configuration
15
from tests.kit.model_zoo import model_zoo, run_fwd_bwd
16

17
PLACEMENT_CONFIGS = [
18
    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.0},  # zero2
19
    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 1.0},  # zero2-offload
20
    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.5},  # zero2-offload-half
21
    {"placement_policy": "static", "shard_param_frac": 1.0},  # zero3
22
    {"placement_policy": "static", "shard_param_frac": 0.5},  # zero3-half
23
    {
24
        "placement_policy": "static",
25
        "shard_param_frac": 1.0,
26
        "offload_optim_frac": 1.0,
27
        "offload_param_frac": 1.0,
28
    },  # zero3-offload-all
29
    {"placement_policy": "auto"},
30
]
31

32
# this model is large enough to slice to chunks
33
TEST_MODELS = ["transformers_gpt_lm"]
34
# these models are too small, all parameters in these models are compacted into one chunk
35
EXAMPLE_MODELS = [
36
    "transformers_bert_for_sequence_classification",
37
    "custom_hanging_param_model",
38
    "custom_nested_model",
39
    "custom_repeated_computed_layers",
40
]
41

42
# bfloat16 cannot represent them exactly
43
BF16_IGNORED_KEYS = [
44
    "masked_bias",
45
]
46

47

48
def check_param(model: GeminiDDP, torch_model: torch.nn.Module, dtype: torch.dtype):
49
    zero_dict = model.state_dict(only_rank_0=False)
50
    torch_dict = torch_model.state_dict()
51

52
    for key, value in torch_dict.items():
53
        # key is 'module.model.PARAMETER', so we truncate it
54
        key = key[7:]
55
        assert key in zero_dict, "{} not in ZeRO dictionary.".format(key)
56
        temp_zero_value = zero_dict[key].to(device=value.device)
57
        if dtype is torch.bfloat16 and any(k in key for k in BF16_IGNORED_KEYS):
58
            continue
59
        rtol, atol = 2e-3, 6e-3
60
        if dtype is torch.bfloat16:
61
            rtol, atol = 4e-3, 8e-3
62
        # debug_print([0], "max range: ", key, torch.max(torch.abs(value - temp_zero_value)))
63
        assert_close(
64
            value.float(),
65
            temp_zero_value.float(),
66
            rtol=rtol,
67
            atol=atol,
68
            msg=lambda s: s + f"\n{key}\n{temp_zero_value.dtype}",
69
        )
70

71

72
@parameterize("placement_config", PLACEMENT_CONFIGS)
73
@parameterize("model_name", TEST_MODELS)
74
@parameterize("mixed_precision", [torch.half, torch.bfloat16])
75
@parameterize("master_weights", [True, False])
76
def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dtype, master_weights: bool):
77
    set_seed(42)
78
    model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
79
        iter(model_zoo.get_sub_registry(model_name).values())
80
    )
81

82
    torch_model = model_builder().cuda()
83
    # apex no master weights leads to nan, so we don't use it
84
    amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=128)
85
    torch_optim = torch.optim.Adam(torch_model.parameters(), lr=1e-3)
86
    torch_model, torch_optim = convert_to_apex_amp(torch_model, torch_optim, amp_config)
87
    torch_model = DDP(torch_model, device_ids=[dist.get_rank()])
88

89
    model = model_builder().cuda()
90

91
    for torch_p, p in zip(torch_model.parameters(), model.parameters()):
92
        p.data.copy_(torch_p.data)
93

94
    world_size = torch.distributed.get_world_size()
95
    config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
96
    config_dict[world_size]["chunk_size"] = 5000
97
    config_dict[world_size]["keep_gathered"] = False
98
    model = GeminiDDP(
99
        model, config_dict, **placement_config, mixed_precision=mixed_precision, master_weights=master_weights
100
    )
101

102
    optimizer = HybridAdam(model.parameters(), lr=1e-3)
103
    zero_optim = GeminiOptimizer(optimizer, model, initial_scale=128)
104

105
    model.eval()
106
    torch_model.eval()
107

108
    set_seed(dist.get_rank() * 3 + 128)
109
    rtol, atol = 4e-2, 4e-2
110
    train_dataloader = iter(DummyDataloader(data_gen_fn))
111
    for i, data in enumerate(train_dataloader):
112
        if i > 2:
113
            break
114
        data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
115
        zero_optim.zero_grad()
116
        torch_optim.zero_grad()
117

118
        torch_loss = run_fwd_bwd(torch_model, data, output_transform_fn, loss_fn, optimizer=torch_optim)
119
        loss = run_fwd_bwd(model, data, output_transform_fn, loss_fn, optimizer=zero_optim)
120
        # as no master weights leads to error accumulation, we don't check the loss
121
        if master_weights:
122
            assert_close(torch_loss.float(), loss.float(), rtol=rtol, atol=atol)
123

124
        zero_optim.step()
125
        torch_optim.step()
126

127
        if master_weights:
128
            check_param(model, torch_model, mixed_precision)
129

130

131
@parameterize("placement_config", [PLACEMENT_CONFIGS[3]])
132
@parameterize("model_name", EXAMPLE_MODELS)
133
@parameterize("mixed_precision", [torch.half])
134
def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.dtype):
135
    set_seed(2008)
136
    model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
137
        iter(model_zoo.get_sub_registry(model_name).values())
138
    )
139

140
    torch_model = model_builder().cuda()
141
    amp_config = dict(opt_level="O2", keep_batchnorm_fp32=False, loss_scale=2)
142
    torch_optim = torch.optim.Adam(torch_model.parameters(), lr=1e-3)
143
    torch_model, torch_optim = convert_to_apex_amp(torch_model, torch_optim, amp_config)
144
    torch_model = DDP(torch_model, device_ids=[dist.get_rank()])
145

146
    model = model_builder().cuda()
147

148
    for torch_p, p in zip(torch_model.parameters(), model.parameters()):
149
        p.data.copy_(torch_p.data)
150

151
    model = GeminiDDP(
152
        model,
153
        chunk_init_device=get_accelerator().get_current_device(),
154
        search_range_m=1,
155
        pin_memory=True,
156
        mixed_precision=mixed_precision,
157
        **placement_config,
158
    )
159
    optimizer = HybridAdam(model.parameters(), lr=1e-3)
160
    zero_optim = GeminiOptimizer(optimizer, model, initial_scale=2)
161

162
    model.eval()
163
    torch_model.eval()
164

165
    set_seed(dist.get_rank() * 3 + 128)
166

167
    train_dataloader = DummyDataloader(data_gen_fn)
168
    for i, data in enumerate(train_dataloader):
169
        if i > 2:
170
            break
171

172
        data = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data.items()}
173

174
        zero_optim.zero_grad()
175
        torch_optim.zero_grad()
176

177
        run_fwd_bwd(torch_model, data, output_transform_fn, loss_fn, optimizer=torch_optim)
178
        run_fwd_bwd(model, data, output_transform_fn, loss_fn, optimizer=zero_optim)
179
        zero_optim.step()
180
        torch_optim.step()
181

182
        check_param(model, torch_model, mixed_precision)
183

184

185
def run_dist(rank, world_size, port):
186
    config = {}
187
    colossalai.launch(config=config, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
188
    exam_model_step()
189
    exam_tiny_example()
190

191

192
@pytest.mark.dist
193
@pytest.mark.parametrize("world_size", [1, 4])
194
@rerun_if_address_is_in_use()
195
def test_optim(world_size):
196
    spawn(run_dist, world_size)
197

198

199
if __name__ == "__main__":
200
    test_optim(1)
201