Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "0efd540dbc5405ada2f57f09d2a376aecad576dc"
Switch branch/tag
Find file Normal viewHistoryPermalink
test_utils.py 5.11 KB
Newer Older
Alex Brooks's avatar
[Bugfix] Fix Crashing When Loading Modules With Batchnorm Stats (#15813)  
Alex Brooks committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
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
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79 
# SPDX-License-Identifier: Apache-2.0

import torch

from vllm.model_executor.models.utils import AutoWeightsLoader


class ModuleWithBatchNorm(torch.nn.Module):

    def __init__(self):
        super().__init__()
        self.bn = torch.nn.BatchNorm1d(2)

    def forward(self, x):
        return self.bn(x)


class ModuleWithNestedBatchNorm(torch.nn.Module):

    def __init__(self):
        super().__init__()
        self.nested_mod = ModuleWithBatchNorm()

    def forward(self, x):
        return self.nested_mod(x)


def test_module_with_batchnorm_can_load():
    """Ensure the auto weight loader can load batchnorm stats."""
    mod = ModuleWithBatchNorm()
    # Run some data through the module with batchnorm
    mod(torch.Tensor([[1, 2], [3, 4]]))

    # Try to load the weights to a new instance
    def weight_generator():
        yield from mod.state_dict().items()

    new_mod = ModuleWithBatchNorm()

    assert not torch.all(new_mod.bn.running_mean == mod.bn.running_mean)
    assert not torch.all(new_mod.bn.running_var == mod.bn.running_var)
    assert new_mod.bn.num_batches_tracked.item() == 0

    loader = AutoWeightsLoader(new_mod)
    loader.load_weights(weight_generator())

    # Ensure the stats are updated
    assert torch.all(new_mod.bn.running_mean == mod.bn.running_mean)
    assert torch.all(new_mod.bn.running_var == mod.bn.running_var)
    assert new_mod.bn.num_batches_tracked.item() == 1


def test_module_with_child_containing_batchnorm_can_autoload():
    """Ensure the auto weight loader can load nested modules batchnorm stats."""
    mod = ModuleWithNestedBatchNorm()
    # Run some data through the module with batchnorm
    mod(torch.Tensor([[1, 2], [3, 4]]))

    # Try to load the weights to a new instance
    def weight_generator():
        yield from mod.state_dict().items()

    new_mod = ModuleWithNestedBatchNorm()

    assert not torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert not torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 0

    loader = AutoWeightsLoader(new_mod)
    loader.load_weights(weight_generator())

    # Ensure the stats are updated
    assert torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 1
Isotr0py's avatar
[Misc] Allow `AutoWeightsLoader` to skip loading weights with specific substr in name (#18358)  
Isotr0py committed
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149 


def test_module_skip_prefix():
    """Ensure the auto weight loader can skip prefix."""
    mod = ModuleWithNestedBatchNorm()
    # Run some data through the module with batchnorm
    mod(torch.Tensor([[1, 2], [3, 4]]))

    # Try to load the weights to a new instance
    def weight_generator():
        # weights needed to be filtered out
        redundant_weights = {
            "prefix.bn.weight": torch.Tensor([1, 2]),
            "prefix.bn.bias": torch.Tensor([3, 4]),
        }
        yield from (mod.state_dict() | redundant_weights).items()

    new_mod = ModuleWithNestedBatchNorm()

    assert not torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert not torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 0

    loader = AutoWeightsLoader(new_mod, skip_prefixes=["prefix."])
    loader.load_weights(weight_generator())

    # Ensure the stats are updated
    assert torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 1


def test_module_skip_substr():
    """Ensure the auto weight loader can skip prefix."""
    mod = ModuleWithNestedBatchNorm()
    # Run some data through the module with batchnorm
    mod(torch.Tensor([[1, 2], [3, 4]]))

    # Try to load the weights to a new instance
    def weight_generator():
        # weights needed to be filtered out
        redundant_weights = {
            "nested_mod.0.substr.weight": torch.Tensor([1, 2]),
            "nested_mod.0.substr.bias": torch.Tensor([3, 4]),
            "nested_mod.substr.weight": torch.Tensor([1, 2]),
            "nested_mod.substr.bias": torch.Tensor([3, 4]),
        }
        yield from (mod.state_dict() | redundant_weights).items()

    new_mod = ModuleWithNestedBatchNorm()

    assert not torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert not torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 0

    loader = AutoWeightsLoader(new_mod, skip_substrs=["substr."])
    loader.load_weights(weight_generator())

    # Ensure the stats are updated
    assert torch.all(
        new_mod.nested_mod.bn.running_mean == mod.nested_mod.bn.running_mean)
    assert torch.all(
        new_mod.nested_mod.bn.running_var == mod.nested_mod.bn.running_var)
    assert new_mod.nested_mod.bn.num_batches_tracked.item() == 1