visual.py 10.2 KB
Newer Older
chenych's avatar
chenych committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
#
# This code is inspired by the HuggingFace's Transformers library.
# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/models/llava/modeling_llava.py
#
# 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.

luopl's avatar
luopl committed
18
19
from dataclasses import dataclass
from typing import TYPE_CHECKING, Dict, List, Optional, Sequence, Set, Tuple, Union
chenych's avatar
chenych committed
20
21

import torch
luopl's avatar
luopl committed
22
import transformers
chenych's avatar
chenych committed
23
24
25
import transformers.models
from transformers.activations import ACT2FN

luopl's avatar
luopl committed
26
from ...extras import logging
chenych's avatar
chenych committed
27
28
29


if TYPE_CHECKING:
luopl's avatar
luopl committed
30
    from transformers import LlavaConfig, PretrainedConfig, PreTrainedModel, ProcessorMixin
chenych's avatar
chenych committed
31

luopl's avatar
luopl committed
32
    from ...hparams import FinetuningArguments, ModelArguments
chenych's avatar
chenych committed
33
34


luopl's avatar
luopl committed
35
36
logger = logging.get_logger(__name__)
transformers_logger = transformers.utils.logging.get_logger(__name__)
chenych's avatar
chenych committed
37
38


luopl's avatar
luopl committed
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
@dataclass
class CompositeModel:
    model_type: str
    projector_key: str
    vision_model_keys: List[str]
    language_model_keys: List[str]

    def get_projector(self, module: "torch.nn.Module") -> "torch.nn.Module":
        for key in self.projector_key.split("."):
            module = getattr(module, key)

        return module


COMPOSITE_MODELS: Dict[str, "CompositeModel"] = {}


def _register_composite_model(
    model_type: str,
    projector_key: Optional[str] = None,
    vision_model_keys: Optional[List[str]] = None,
    language_model_keys: Optional[List[str]] = None,
):
    projector_key = projector_key or "multi_modal_projector"
    vision_model_keys = vision_model_keys or ["vision_tower"]
    language_model_keys = language_model_keys or ["language_model"]
    COMPOSITE_MODELS[model_type] = CompositeModel(
        model_type=model_type,
        projector_key=projector_key,
        vision_model_keys=vision_model_keys,
        language_model_keys=language_model_keys,
    )


chenych's avatar
chenych committed
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
class LlavaMultiModalProjectorForYiVL(torch.nn.Module):
    def __init__(self, config: "LlavaConfig") -> None:
        super().__init__()

        self.config = config
        if config is None:
            return

        self.linear_1 = torch.nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.linear_2 = torch.nn.LayerNorm(config.text_config.hidden_size, bias=True)
        self.linear_3 = torch.nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
        self.linear_4 = torch.nn.LayerNorm(config.text_config.hidden_size, bias=True)
        self.act = ACT2FN[config.projector_hidden_act]

    def forward(self, image_features: "torch.Tensor") -> "torch.Tensor":
        hidden_states = self.linear_1(image_features)
        hidden_states = self.linear_2(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.linear_3(hidden_states)
        hidden_states = self.linear_4(hidden_states)
        if hidden_states.dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.linear_1.weight.dtype

            transformers_logger.warning_once("The hidden states seems to be silently casted in float32.")
            hidden_states = hidden_states.to(target_dtype)

        return hidden_states


class LlavaMultiModalProjectorForYiVLForVLLM(LlavaMultiModalProjectorForYiVL):
    def __init__(self, vision_hidden_size: int, text_hidden_size: int, projector_hidden_act: str) -> None:
        super().__init__(config=None)

        self.linear_1 = torch.nn.Linear(vision_hidden_size, text_hidden_size, bias=True)
        self.linear_2 = torch.nn.LayerNorm(text_hidden_size, bias=True)
        self.linear_3 = torch.nn.Linear(text_hidden_size, text_hidden_size, bias=True)
        self.linear_4 = torch.nn.LayerNorm(text_hidden_size, bias=True)
        self.act = ACT2FN[projector_hidden_act]


luopl's avatar
luopl committed
118
119
120
121
122
def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArguments") -> None:
    r"""
    Casts projector output to half precision for fine-tuning quantized VLMs.
    """

chenych's avatar
chenych committed
123
124
125
126
127
    def _mm_projector_forward_post_hook(
        module: "torch.nn.Module", args: Tuple["torch.Tensor"], output: "torch.Tensor"
    ) -> "torch.Tensor":
        return output.to(model_args.compute_dtype)

luopl's avatar
luopl committed
128
129
    if getattr(model, "quantization_method", None):
        model_type = getattr(model.config, "model_type", None)
luopl's avatar
luopl committed
130
131
        if model_type in COMPOSITE_MODELS:
            mm_projector = COMPOSITE_MODELS[model_type].get_projector(model)
luopl's avatar
luopl committed
132
133
134
        else:
            return

luopl's avatar
luopl committed
135
        logger.info_rank0(f"Casting multimodal projector outputs in {model_args.compute_dtype}.")
chenych's avatar
chenych committed
136
137
138
139
        mm_projector.register_forward_hook(_mm_projector_forward_post_hook)


def configure_visual_model(config: "PretrainedConfig") -> None:
luopl's avatar
luopl committed
140
141
142
    r"""
    Patches VLMs before loading them.
    """
luopl's avatar
luopl committed
143
144
    if getattr(config, "text_config", None) and not getattr(config, "hidden_size", None):
        # required for ds zero3 and valuehead models
chenych's avatar
chenych committed
145
146
147
        setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))

    if getattr(config, "is_yi_vl_derived_model", None):
luopl's avatar
luopl committed
148
        logger.info_rank0("Detected Yi-VL model, applying projector patch.")
chenych's avatar
chenych committed
149
        transformers.models.llava.modeling_llava.LlavaMultiModalProjector = LlavaMultiModalProjectorForYiVL
luopl's avatar
luopl committed
150
151
152
153
154
155
156
157


def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "FinetuningArguments") -> Set[str]:
    r"""
    Freezes vision tower and language model for VLM full/freeze tuning.
    """
    model_type = getattr(config, "model_type", None)
    forbidden_modules = set()
luopl's avatar
luopl committed
158
    if model_type in COMPOSITE_MODELS:
luopl's avatar
luopl committed
159
        if finetuning_args.freeze_vision_tower:
luopl's avatar
luopl committed
160
161
162
            vision_model_keys = COMPOSITE_MODELS[model_type].vision_model_keys
            logger.info_rank0(f"Set vision model not trainable: {vision_model_keys}.")
            forbidden_modules.update(vision_model_keys)
luopl's avatar
luopl committed
163

luopl's avatar
luopl committed
164
165
166
167
        if finetuning_args.freeze_multi_modal_projector:
            projector_key = COMPOSITE_MODELS[model_type].projector_key
            logger.info_rank0(f"Set multi model projector not trainable: {projector_key}.")
            forbidden_modules.add(projector_key)
luopl's avatar
luopl committed
168
169

        if finetuning_args.train_mm_proj_only:
luopl's avatar
luopl committed
170
171
172
            language_model_keys = COMPOSITE_MODELS[model_type].language_model_keys
            logger.info_rank0(f"Set language model not trainable: {language_model_keys}.")
            forbidden_modules.update(language_model_keys)
luopl's avatar
luopl committed
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193

    return forbidden_modules


def get_image_seqlen(config: "PretrainedConfig") -> int:
    r"""
    Computes the number of special tokens per image.
    """
    model_type = getattr(config, "model_type", None)
    if model_type == "llava":
        image_seqlen = (config.vision_config.image_size // config.vision_config.patch_size) ** 2
        if getattr(config, "vision_feature_select_strategy", "default") == "full":  # add [CLS] token
            image_seqlen += 1
    elif model_type == "paligemma":
        image_seqlen = config.vision_config.num_image_tokens
    else:
        image_seqlen = -1

    return image_seqlen


luopl's avatar
luopl committed
194
def get_patch_size(config: "PretrainedConfig", processor: "ProcessorMixin") -> int:
luopl's avatar
luopl committed
195
196
197
    r"""
    Computes the patch size of the vit.
    """
luopl's avatar
luopl committed
198
    patch_size = getattr(config.vision_config, "patch_size", getattr(processor, "patch_size", -1))
luopl's avatar
luopl committed
199
200
201
    return patch_size


luopl's avatar
luopl committed
202
def get_vision_feature_select_strategy(config: "PretrainedConfig", processor: "ProcessorMixin") -> int:
luopl's avatar
luopl committed
203
204
205
    r"""
    Get the vision_feature_select_strategy.
    """
luopl's avatar
luopl committed
206
207
208
    vision_feature_select_strategy = getattr(
        config, "vision_feature_select_strategy", getattr(processor, "vision_feature_select_strategy", "default")
    )
luopl's avatar
luopl committed
209
210
211
212
213
214
215
216
217
218
    return vision_feature_select_strategy


def patch_target_modules(
    config: "PretrainedConfig", finetuning_args: "FinetuningArguments", target_modules: Sequence[str]
) -> Union[str, List[str]]:
    r"""
    Freezes vision tower for VLM LoRA tuning.
    """
    model_type = getattr(config, "model_type", None)
luopl's avatar
luopl committed
219
    vit_model_type = getattr(getattr(config, "vision_config", None), "model_type", None)
luopl's avatar
luopl committed
220
    if finetuning_args.freeze_vision_tower:
luopl's avatar
luopl committed
221
222
223
224
225
226
        if model_type in COMPOSITE_MODELS:
            vision_model_keys = COMPOSITE_MODELS[model_type].vision_model_keys
            logger.info_rank0(f"Set vision model not trainable: {vision_model_keys}.")
            vision_model_keys = "|".join(vision_model_keys)
            target_modules = "|".join(target_modules)
            return f"^(?!.*{vision_model_keys}).*(?:{target_modules}).*"
luopl's avatar
luopl committed
227
228
229
        else:
            return target_modules
    else:
luopl's avatar
luopl committed
230
        if model_type == "qwen2_vl":  # avoid attaching lora to Conv3D layer
luopl's avatar
luopl committed
231
            return "^(?!.*patch_embed).*(?:{}).*".format("|".join(target_modules))
luopl's avatar
luopl committed
232
        elif vit_model_type == "pixtral":
luopl's avatar
luopl committed
233
            return "^(?!.*patch_conv).*(?:{}).*".format("|".join(target_modules))
luopl's avatar
luopl committed
234
235
        else:
            return target_modules
luopl's avatar
luopl committed
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288


_register_composite_model(
    model_type="llava",
)


_register_composite_model(
    model_type="llava_next",
)


_register_composite_model(
    model_type="llava_next_video",
)


_register_composite_model(
    model_type="minicpmv",
    vision_model_keys=["vpm"],
    language_model_keys=["llm"],
)


_register_composite_model(
    model_type="minicpmo",
    vision_model_keys=["vpm", "apm", "resampler", "tts"],
    language_model_keys=["llm"],
)


_register_composite_model(
    model_type="paligemma",
)


_register_composite_model(
    model_type="video_llava",
)


_register_composite_model(
    model_type="mllama",
    vision_model_keys=["vision_model"],
)


_register_composite_model(
    model_type="qwen2_vl",
    projector_key="visual.merger",
    vision_model_keys=["visual.patch_embed", "visual.blocks"],
    language_model_keys=["model", "lm_head"],
)