File size: 6,289 Bytes
f0395ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
150
151
152
153
154
155
156
157
158
159
160
161
# Copyright (c) 2025 Hansheng Chen

import importlib
import torch
import torch.nn as nn
import torch.distributed as dist

try:
    from torch.distributed.fsdp import fully_shard, MixedPrecisionPolicy
except:
    pass
from mmcv.parallel.scatter_gather import scatter_kwargs
from mmcv.parallel import MODULE_WRAPPERS


def get_module_object(path):
    module_path, attribute = path.rsplit('.', 1)
    module = importlib.import_module(module_path)
    return getattr(module, attribute)


@MODULE_WRAPPERS.register_module()
class FSDP2Wrapper(nn.Module):

    def __init__(
            self,
            module,
            wrap_frozen_modules=False,
            ignore_frozen_parameters=False,
            param_dtype='bfloat16',
            reduce_dtype='float32',
            fsdp_modules=None,
            exclude_keys=(),
            hybrid_sharding=True,
            **kwargs):
        super().__init__()
        self.module = module
        fsdp_kwargs = kwargs
        self.param_dtype = getattr(torch, param_dtype)
        self.reduce_dtype = getattr(torch, reduce_dtype)
        if hybrid_sharding:
            global_world_size = dist.get_world_size()
            num_devices_per_node = torch.cuda.device_count()
            mesh = dist.init_device_mesh(
                'cuda',
                (global_world_size // num_devices_per_node, num_devices_per_node),
                mesh_dim_names=('replicate', 'shard'))
            fsdp_kwargs.update(mesh=mesh)
        if fsdp_modules is not None:
            assert isinstance(fsdp_modules, (list, tuple))
            fsdp_modules = tuple([get_module_object(m) for m in fsdp_modules])
        self.to_fsdp(
            wrap_frozen_modules,
            ignore_frozen_parameters,
            exclude_keys,
            fsdp_modules=fsdp_modules,
            **fsdp_kwargs)

    def to_fsdp(self,
                wrap_frozen_modules=False,
                ignore_frozen_parameters=False,
                exclude_keys=(),
                fsdp_modules=(),
                **kwargs):
        for name, module in self.module._modules.items():
            if name in exclude_keys or next(module.parameters(), None) is None:
                module = module.cuda()
            elif all(not p.requires_grad for p in module.parameters()):
                if wrap_frozen_modules:
                    for submodule in module.modules():
                        if isinstance(submodule, fsdp_modules):
                            fsdp_kwargs = kwargs.copy()
                            fsdp_kwargs.update(
                                mp_policy=MixedPrecisionPolicy(
                                    param_dtype=self.param_dtype,
                                    reduce_dtype=self.reduce_dtype))
                            fully_shard(submodule, **fsdp_kwargs)
                    fsdp_kwargs = kwargs.copy()
                    fsdp_kwargs.update(
                        mp_policy=MixedPrecisionPolicy(
                            param_dtype=self.param_dtype,
                            reduce_dtype=self.reduce_dtype,
                            cast_forward_inputs=False))
                    fully_shard(module, **fsdp_kwargs)
                else:
                    module = module.cuda()
            else:
                if ignore_frozen_parameters:
                    ignored_params = []
                    for p in module.parameters():
                        if not p.requires_grad:
                            p.data = p.data.cuda()
                            ignored_params.append(p)
                else:
                    ignored_params = None
                for submodule in module.modules():
                    if isinstance(submodule, fsdp_modules):
                        fsdp_kwargs = kwargs.copy()
                        fsdp_kwargs.update(
                            mp_policy=MixedPrecisionPolicy(
                                param_dtype=self.param_dtype,
                                reduce_dtype=self.reduce_dtype))
                        if ignored_params is not None:  # requires torch >= 2.7
                            fsdp_kwargs.update(ignored_params=ignored_params)
                        fully_shard(submodule, **fsdp_kwargs)
                fsdp_kwargs = kwargs.copy()
                fsdp_kwargs.update(
                    mp_policy=MixedPrecisionPolicy(
                        param_dtype=self.param_dtype,
                        reduce_dtype=self.reduce_dtype,
                        cast_forward_inputs=False))
                if ignored_params is not None:  # requires torch >= 2.7
                    fsdp_kwargs.update(ignored_params=ignored_params)
                fully_shard(module, **fsdp_kwargs)
            self.module._modules[name] = module

    def scatter(self, inputs, kwargs, device_ids):
        """Scatter function.

        Args:
            inputs (Tensor): Input Tensor.
            kwargs (dict): Args for
                ``mmcv.parallel.scatter_gather.scatter_kwargs``.
            device_ids (int): Device id.
        """
        return scatter_kwargs(inputs, kwargs, device_ids)

    def forward(self, *inputs, **kwargs):
        """Forward function.

        Args:
            inputs (tuple): Input data.
            kwargs (dict): Args for
                ``mmcv.parallel.scatter_gather.scatter_kwargs``.
        """
        inputs, kwargs = self.scatter(inputs, kwargs, [torch.cuda.current_device()])
        return self.module(*inputs[0], **kwargs[0])

    def train_step(self, *inputs, **kwargs):
        """Train step function.

        Args:
            inputs (Tensor): Input Tensor.
            kwargs (dict): Args for
                ``mmcv.parallel.scatter_gather.scatter_kwargs``.
        """
        inputs, kwargs = self.scatter(inputs, kwargs, [torch.cuda.current_device()])
        output = self.module.train_step(*inputs[0], **kwargs[0])
        return output

    def val_step(self, *inputs, **kwargs):
        """Validation step function.

        Args:
            inputs (tuple): Input data.
            kwargs (dict): Args for ``scatter_kwargs``.
        """
        inputs, kwargs = self.scatter(inputs, kwargs, [torch.cuda.current_device()])
        output = self.module.val_step(*inputs[0], **kwargs[0])
        return output