File size: 7,640 Bytes
76f9669
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from abc import ABC, abstractmethod
from typing import List, Optional

import torch
import torch.nn.utils.parametrize as P
import tqdm
from compressed_tensors.modeling.attention import (
    initialize_hooked_attention,
    register_query_hook,
)
from compressed_tensors.modeling.kvcache import (
    initialize_hooked_kv_cache,
    register_key_hook,
)
from compressed_tensors.registry.registry import RegistryMixin, T
from compressed_tensors.transform import (
    TransformArgs,
    TransformLocation,
    TransformScheme,
)
from compressed_tensors.utils import (
    align_module_device,
    delete_offload_module,
    has_offloaded_params,
    match_named_modules,
    patch_attr,
    register_offload_module,
    update_offload_parameter,
)
from compressed_tensors.utils.internal import InternalModule
from torch import Tensor
from torch.nn import Module, Parameter
from transformers import PreTrainedModel


__all__ = ["TransformFactory", "TransformBase"]


class TransformFactory(RegistryMixin, ABC):
    """
    Abstract factory base used to create and apply transforms to a model

    :param name: name associated with transform scheme
    :param scheme: transform scheme which defines how transforms should be created
    :param seed: random seed used to transform weight randomization
    """

    transforms: List["TransformBase"]

    def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
        self.name = name
        self.scheme = scheme
        self.generator = torch.Generator()
        if seed is not None:
            self.generator.manual_seed(seed)

    @classmethod
    def from_scheme(cls: type[T], scheme: TransformScheme, **kwargs) -> T:
        """
        Create a transform factory from a scheme

        :param scheme: defines how transforms should be created
        :param kwargs: TransformFactory constructor arguments
        :return: subclass of `TransformFactory` corresponding to the scheme type
        """
        constructor = cls.get_value_from_registry(name=scheme.type)
        return constructor(scheme=scheme, **kwargs)

    @abstractmethod
    def create_transform(self, module: Module, args: TransformArgs) -> "TransformBase":
        """
        Abstract method which defines how a transform should be created. May utilize
        caching to maximize shared memory

        :param module: parent module that transform will be applied to
        :param args: defines how the transform will be applied to the module
        :return: instance of TransformBase
        """
        raise NotImplementedError()

    def apply_to_model(self, model: Module, use_tqdm=True):
        """
        Create transforms and apply them to the model

        :param model: module to apply transforms to
        """
        modules_args = [
            (module, arg)
            for arg in self.scheme.apply
            for _, module in match_named_modules(model, arg.targets, arg.ignore)
        ]

        desc = f"Applying {self.name} transforms"
        for module, arg in tqdm.tqdm(modules_args, desc=desc, disable=(not use_tqdm)):
            self._apply_to_module(model, module, arg)

    def _apply_to_module(self, model: Module, module: Module, args: TransformArgs):
        """
        Create transforms and apply them to the module

        :param model: model which module belongs to
        :param module: target module to apply transforms to
        :param args: defines how the transform will be applied to the target module
        """
        if has_offloaded_params(module):
            if module._hf_hook.place_submodules:
                raise NotImplementedError(
                    "Applying transforms to offloaded submodules with "
                    "`place_submodules=True` is not supported"
                )

        # create transform as submodule
        transform_name = f"{self.name}_{args.location}"
        transform = self.create_transform(module, args)
        register_offload_module(module, transform_name, transform)

        # register input transformation hook
        if args.location == TransformLocation.INPUT:

            def input_hook(_, args):
                input = args[0]
                return transform(input)

            module.register_forward_pre_hook(input_hook, prepend=True)

        # eagerly apply transformation to weight
        elif args.location in (
            TransformLocation.WEIGHT_INPUT,
            TransformLocation.WEIGHT_OUTPUT,
        ):
            # fuse transform into weight
            assert hasattr(module, "weight")
            with torch.no_grad(), align_module_device(module):
                update_offload_parameter(module, "weight", transform(module.weight))

            if self.scheme.requires_grad:
                # for training, the weight changes with every forward pass
                # so we can leverage parametrization to propagate the gradient
                if has_offloaded_params(module):
                    raise ValueError("Offloaded training is not supported")
                P.register_parametrization(module, "weight", transform)

            else:
                # transform is no longer needed (unfusing is not supported)
                delete_offload_module(module, transform_name)

        # register output transformation hook
        elif args.location == TransformLocation.OUTPUT:

            def output_hook(_, _input, output):
                return transform(output)

            module.register_forward_hook(output_hook)

        # register query hook to attention
        elif args.location == TransformLocation.Q_ATTN:
            if not isinstance(model, PreTrainedModel):
                raise ValueError(f"Cannot hook attention of model: {model}")

            def query_hook(_, query_states):
                return transform(query_states)

            initialize_hooked_attention(model, module)
            register_query_hook(module, query_hook)

        # register key hook to kvcache
        elif args.location == TransformLocation.K_CACHE:
            if not isinstance(model, PreTrainedModel):
                raise ValueError(f"Cannot hook attention of model: {model}")

            def key_hook(_, key_states):
                return transform(key_states)

            initialize_hooked_kv_cache(model, module)
            register_key_hook(module, key_hook)

        else:
            raise NotImplementedError()


class TransformBase(InternalModule, ABC):
    """
    Represents the application of a transform accord to TransformArgs
    """

    args: TransformArgs
    weight: Parameter
    _dynamic_tied_weights_keys: List[str] = ["weight"]

    @abstractmethod
    def forward(self, value: Tensor) -> Tensor:
        raise NotImplementedError()

    def right_inverse(self, value: Tensor) -> Tensor:
        with patch_attr(self.args, "inverse", not self.args.inverse):
            return self.forward(value)

    def __repr__(self):
        return f"{self.__class__.__name__}(inverse={self.args.inverse})"