interns1_vit.py 14.7 KB
Newer Older
Lyu Han's avatar
Lyu Han committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

# adapted from https://huggingface.co/OpenGVLab/InternVL2-4B/blob/main/modeling_intern_vit.py
# --------------------------------------------------------
# InternVL
# Copyright (c) 2023 OpenGVLab
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------
from collections.abc import Iterable
from typing import Optional

import torch
import torch.nn as nn
from transformers import PretrainedConfig
from transformers.utils import torch_int

18
from vllm.attention.layer import MultiHeadAttention
Lyu Han's avatar
Lyu Han committed
19
20
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.layernorm import RMSNorm
21
from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear
Lyu Han's avatar
Lyu Han committed
22
23
24
25
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader

NORM2FN = {
26
27
    "rms_norm": RMSNorm,
    "layer_norm": nn.LayerNorm,
Lyu Han's avatar
Lyu Han committed
28
29
30
31
32
33
34
35
36
}


class InternS1VisionPatchEmbeddings(nn.Module):
    def __init__(self, config):
        super().__init__()
        image_size, patch_size = config.image_size, config.patch_size
        num_channels, hidden_size = config.num_channels, config.hidden_size

37
38
39
40
        num_patches = (image_size[1] // patch_size[1]) * (
            image_size[0] // patch_size[0]
        )
        patch_shape = (image_size[0] // patch_size[0], image_size[1] // patch_size[1])
Lyu Han's avatar
Lyu Han committed
41
42
43
44
45
46
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_channels = num_channels
        self.num_patches = num_patches
        self.patch_shape = patch_shape

47
48
49
        self.projection = nn.Conv2d(
            num_channels, hidden_size, kernel_size=patch_size, stride=patch_size
        )
Lyu Han's avatar
Lyu Han committed
50
51
52
53
54
55

    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
        batch_size, num_channels, height, width = pixel_values.shape
        if num_channels != self.num_channels:
            raise ValueError(
                "Make sure that the channel dimension of the pixel values "
56
57
                "match with the one set in the configuration."
            )
Lyu Han's avatar
Lyu Han committed
58

59
        embeddings = self.projection(pixel_values.to(self.projection.weight.dtype))
Lyu Han's avatar
Lyu Han committed
60
61
62
63
64
65
66
67
68
69
70
71
        patch_height, patch_width = embeddings.shape[2], embeddings.shape[3]
        embeddings = embeddings.flatten(2).transpose(1, 2)

        return embeddings, (patch_height, patch_width)


class InternS1VisionEmbeddings(nn.Module):
    def __init__(self, config: PretrainedConfig):
        super().__init__()
        self.config = config
        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        if config.use_mask_token:
72
            self.mask_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
Lyu Han's avatar
Lyu Han committed
73
74
75
76
        else:
            self.mask_token = None
        self.patch_embeddings = InternS1VisionPatchEmbeddings(config)
        self.patch_size = config.patch_size
77
78
79
80
81
        self.image_size = (
            config.image_size
            if isinstance(config.image_size, Iterable)
            else (config.image_size, config.image_size)
        )
Lyu Han's avatar
Lyu Han committed
82
83
84
        num_patches = self.patch_embeddings.num_patches
        if config.use_absolute_position_embeddings:
            self.position_embeddings = nn.Parameter(
85
86
                torch.zeros(1, num_patches + 1, config.hidden_size)
            )
Lyu Han's avatar
Lyu Han committed
87
88
89
        else:
            self.position_embeddings = None

90
91
92
    def interpolate_pos_encoding(
        self, embeddings: torch.Tensor, height: int, width: int
    ) -> torch.Tensor:
Lyu Han's avatar
Lyu Han committed
93
94
95
96
97
98
99
100
101
102
103
104
105
106
        """
        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
        images. This method is also adapted to support torch.jit tracing.

        Adapted from:
        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
        """  # noqa: E501

        num_patches = embeddings.shape[1] - 1
        num_positions = self.position_embeddings.shape[1] - 1

        # always interpolate when tracing to ensure the exported model
        # works for dynamic input shapes
107
108
109
110
111
        if (
            not torch.jit.is_tracing()
            and num_patches == num_positions
            and height == width
        ):
Lyu Han's avatar
Lyu Han committed
112
113
114
115
116
117
118
119
120
121
122
            return self.position_embeddings

        class_pos_embed = self.position_embeddings[:, :1]
        patch_pos_embed = self.position_embeddings[:, 1:]

        dim = embeddings.shape[-1]

        new_height = height // self.patch_size[0]
        new_width = width // self.patch_size[1]

        sqrt_num_positions = torch_int(num_positions**0.5)
123
124
125
        patch_pos_embed = patch_pos_embed.reshape(
            1, sqrt_num_positions, sqrt_num_positions, dim
        )
Lyu Han's avatar
Lyu Han committed
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)

        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed,
            size=(new_height, new_width),
            mode="bicubic",
            align_corners=False,
        )

        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)

        return torch.cat((class_pos_embed, patch_pos_embed), dim=1)

    def forward(
        self,
        pixel_values: torch.Tensor,
        bool_masked_pos: Optional[torch.BoolTensor] = None,
    ) -> torch.Tensor:
        _, _, height, width = pixel_values.shape
145
        embeddings, (patch_height, patch_width) = self.patch_embeddings(pixel_values)
Lyu Han's avatar
Lyu Han committed
146
147
148
149
150
151
152
153
154
155
156
157
158
        batch_size, seq_len, _ = embeddings.size()

        if bool_masked_pos is not None:
            mask_tokens = self.mask_token.expand(batch_size, seq_len, -1)
            # replace the masked visual tokens by mask_tokens
            w = bool_masked_pos.unsqueeze(-1).type_as(mask_tokens)
            embeddings = embeddings * (1 - w) + mask_tokens * w

        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
        embeddings = torch.cat((cls_tokens, embeddings), dim=1)

        if self.position_embeddings is not None:
            embeddings = embeddings + self.interpolate_pos_encoding(
159
160
                embeddings, height, width
            )
Lyu Han's avatar
Lyu Han committed
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181

        return embeddings, (patch_height, patch_width)


class InternSdpaAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        config: PretrainedConfig,
        *,
        num_dummy_heads: int = 0,
    ) -> None:
        super().__init__()

        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
182
183
184
185
                f"embed_dim must be divisible by num_heads "
                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads})."
            )
Lyu Han's avatar
Lyu Han committed
186
187
188
189
190
191

        # Additional dummy heads are used to enable TP for common GPU counts.
        self.dummy_dim = (num_dummy_heads + self.num_heads) * self.head_dim

        self.scale = self.head_dim**-0.5

192
193
194
195
196
197
198
199
200
        self.q_proj = nn.Linear(
            self.embed_dim, self.num_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            self.embed_dim, self.num_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            self.embed_dim, self.num_heads * self.head_dim, bias=config.attention_bias
        )
Lyu Han's avatar
Lyu Han committed
201
202
203

        self.qk_normalization = config.use_qk_norm
        if self.qk_normalization:
204
205
206
207
208
209
210
211
212
213
            self.q_norm = RMSNorm(
                self.dummy_dim,
                eps=config.layer_norm_eps,
                var_hidden_size=self.embed_dim,
            )
            self.k_norm = RMSNorm(
                self.dummy_dim,
                eps=config.layer_norm_eps,
                var_hidden_size=self.embed_dim,
            )
Lyu Han's avatar
Lyu Han committed
214
215
216

        self.projection_layer = nn.Linear(self.dummy_dim, self.embed_dim)

217
        # Use unified MultiHeadAttention with automatic backend selection
218
        self.attn = MultiHeadAttention(self.num_heads, self.head_dim, self.scale)
219

Lyu Han's avatar
Lyu Han committed
220
221
222
223
224
225
226
227
228
229
230
231
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, N, C = x.shape

        q = self.q_proj(x)
        k = self.k_proj(x)
        v = self.v_proj(x)

        if self.qk_normalization:
            B_, N_, H_, D_ = q.shape
            q = self.q_norm(q.flatten(-2, -1)).view(B_, N_, H_, D_)
            k = self.k_norm(k.flatten(-2, -1)).view(B_, N_, H_, D_)

232
233
        # Use unified MultiHeadAttention with automatic backend selection
        x = self.attn(q, k, v)
Lyu Han's avatar
Lyu Han committed
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249

        x = self.projection_layer(x)
        return x


class InternS1VisionMLP(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.config = config
        self.activation_fn = get_act_fn(config.hidden_act)
250
251
252
253
254
255
256
257
258
259
260
261
262
263
        self.fc1 = ColumnParallelLinear(
            config.hidden_size,
            config.intermediate_size,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.fc1",
        )
        self.fc2 = RowParallelLinear(
            config.intermediate_size,
            config.hidden_size,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.fc2",
        )
Lyu Han's avatar
Lyu Han committed
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)

        return hidden_states


class InternS1VisionLayer(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_dummy_heads: int = 0,
        prefix: str = "",
    ) -> None:
        super().__init__()

284
285
286
287
288
289
        self.attention = self._init_attn(
            config,
            quant_config,
            num_dummy_heads=num_dummy_heads,
            prefix=f"{prefix}.attention",
        )
Lyu Han's avatar
Lyu Han committed
290

291
292
293
        self.mlp = InternS1VisionMLP(
            config, quant_config=quant_config, prefix=f"{prefix}.mlp"
        )
Lyu Han's avatar
Lyu Han committed
294
        self.layernorm_before = NORM2FN[config.norm_type](
295
296
            config.hidden_size, eps=config.layer_norm_eps
        )
Lyu Han's avatar
Lyu Han committed
297
        self.layernorm_after = NORM2FN[config.norm_type](
298
299
            config.hidden_size, eps=config.layer_norm_eps
        )
Lyu Han's avatar
Lyu Han committed
300
301

        init_values = config.layer_scale_init_value
302
303
304
305
306
307
        self.lambda_1 = nn.Parameter(
            init_values * torch.ones(config.hidden_size), requires_grad=True
        )
        self.lambda_2 = nn.Parameter(
            init_values * torch.ones(config.hidden_size), requires_grad=True
        )
Lyu Han's avatar
Lyu Han committed
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322

    def _init_attn(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
        *,
        num_dummy_heads: int,
        prefix: str = "",
    ):
        return InternSdpaAttention(config, num_dummy_heads=num_dummy_heads)

    def forward(
        self,
        hidden_states: torch.Tensor,
    ):
323
324
325
326
        hidden_states = (
            hidden_states
            + self.attention(self.layernorm_before(hidden_states)) * self.lambda_1
        )
Lyu Han's avatar
Lyu Han committed
327

328
329
330
331
        hidden_states = (
            hidden_states
            + self.mlp(self.layernorm_after(hidden_states)) * self.lambda_2
        )
Lyu Han's avatar
Lyu Han committed
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354

        return hidden_states


class InternS1VisionEncoder(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        num_dummy_heads: int = 0,
        prefix: str = "",
    ):
        super().__init__()

        self.config = config

        if num_hidden_layers_override is None:
            num_hidden_layers = config.num_hidden_layers
        else:
            num_hidden_layers = num_hidden_layers_override

355
356
357
358
359
360
361
362
363
364
365
        self.layer = nn.ModuleList(
            [
                InternS1VisionLayer(
                    config,
                    quant_config,
                    num_dummy_heads=num_dummy_heads,
                    prefix=f"{prefix}.layer.{layer_idx}",
                )
                for layer_idx in range(num_hidden_layers)
            ]
        )
Lyu Han's avatar
Lyu Han committed
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395

    def forward(self, inputs_embeds: torch.Tensor):
        hidden_states = inputs_embeds
        for encoder_layer in self.layer:
            hidden_states = encoder_layer(hidden_states)

        return hidden_states


class InternS1VisionModel(nn.Module):
    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        num_dummy_heads: int = 0,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.config = config

        self.embeddings = InternS1VisionEmbeddings(config)
        self.encoder = InternS1VisionEncoder(
            config=config,
            num_hidden_layers_override=num_hidden_layers_override,
            num_dummy_heads=num_dummy_heads,
            prefix=f"{prefix}.encoder",
        )
396
397
398
399
400
        self.layernorm = (
            nn.Identity()
            if config.use_mean_pooling
            else nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        )
Lyu Han's avatar
Lyu Han committed
401
402
403
404
405
406
407
408
409
410

    def get_input_embeddings(self):
        return self.embeddings.patch_embeddings

    def forward(
        self,
        pixel_values: Optional[torch.Tensor] = None,
        pixel_embeds: Optional[torch.Tensor] = None,
    ) -> torch.FloatTensor:
        if pixel_values is None and pixel_embeds is None:
411
            raise ValueError("You have to specify pixel_values or pixel_embeds")
Lyu Han's avatar
Lyu Han committed
412
413
414
415
416
417
418

        if pixel_embeds is not None:
            hidden_states = pixel_embeds
        elif pixel_values is not None:
            if pixel_values.ndim == 4:
                hidden_states, _ = self.embeddings(pixel_values)
            else:
419
                raise ValueError(f"wrong pixel_values size: {pixel_values.shape}")
Lyu Han's avatar
Lyu Han committed
420
421
422
423
424
425

        encoder_outputs = self.encoder(inputs_embeds=hidden_states)
        encoder_outputs = self.layernorm(encoder_outputs)

        return encoder_outputs

426
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
Lyu Han's avatar
Lyu Han committed
427
428
429
430
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            param = params_dict[name]
431
            weight_loader = getattr(param, "weight_loader", default_weight_loader)
Lyu Han's avatar
Lyu Han committed
432
433
434
            weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params