phi3v.py 28.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# Copyright 2024 The vLLM team.
# Copyright 2024 Microsoft and the HuggingFace Inc. team. 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.
15
import itertools
16
import re
17
from functools import cached_property, lru_cache
18
19
from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional,
                    Tuple, TypedDict, Union)
20

21
import numpy as np
22
23
import torch
import torch.nn as nn
24
from PIL import Image
25
from transformers import CLIPVisionConfig, PretrainedConfig
26
27

from vllm.attention import AttentionMetadata
28
from vllm.config import ModelConfig, VllmConfig
29
30
from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, DummyData,
                         InputContext, token_inputs)
31
from vllm.logger import init_logger
32
from vllm.model_executor.layers.pooler import Pooler, PoolingType
33
from vllm.model_executor.layers.quantization import QuantizationConfig
Joe Runde's avatar
Joe Runde committed
34
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
35
36
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
37
from vllm.model_executor.models.clip import CLIPVisionModel
38
from vllm.model_executor.models.llama import LlamaForCausalLM
39
from vllm.model_executor.pooling_metadata import PoolingMetadata
40
from vllm.model_executor.sampling_metadata import SamplingMetadata
41
from vllm.multimodal import MULTIMODAL_REGISTRY
42
from vllm.multimodal.utils import cached_get_tokenizer, repeat_and_pad_token
43
from vllm.sequence import IntermediateTensors, PoolerOutput
44
from vllm.utils import is_list_of
45

46
from .clip import dummy_image_for_clip, dummy_seq_data_for_clip
47
from .interfaces import SupportsMultiModal, SupportsPP
48
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
49
                    merge_multimodal_embeddings)
50

51
52
logger = init_logger(__name__)

53
54
55
# Cannot find the following 2 numbers from hf config.
_IMAGE_TOKEN_ID = 32044

56
57
58
59
# Result in the max possible feature size (h:w = 16:1)
MAX_IMAGE_FEATURE_SIZE_HEIGHT = 8000
MAX_IMAGE_FEATURE_SIZE_WIDTH = 50

60
61
62
63
64
65
66
67
68
69
70
71
CLIP_VIT_LARGE_PATCH14_336_CONFIG = CLIPVisionConfig(dropout=0.0,
                                                     hidden_act="quick_gelu",
                                                     hidden_size=1024,
                                                     image_size=336,
                                                     intermediate_size=4096,
                                                     num_attention_heads=16,
                                                     num_channels=3,
                                                     num_hidden_layers=24,
                                                     patch_size=14,
                                                     projection_dim=768)


72
def _init_img_processor(hf_config: PretrainedConfig,
73
74
                        quant_config: Optional[QuantizationConfig],
                        prefix: str = "") -> CLIPVisionModel:
75
76
77
78
79
80
81
82
83
84
85
    clip_config = CLIP_VIT_LARGE_PATCH14_336_CONFIG
    layer_idx = hf_config.img_processor.get('layer_idx', -2)

    # Initialize the CLIP only up to the required feature layer
    if layer_idx < 0:
        num_hidden_layers = clip_config.num_hidden_layers + \
            layer_idx + 1
    else:
        num_hidden_layers = layer_idx + 1

    img_processor = CLIPVisionModel(
86
87
88
        clip_config,
        quant_config,
        num_hidden_layers_override=num_hidden_layers,
89
        prefix=prefix,
90
    )
91
92
93
94

    return img_processor


95
96
97
98
class Phi3VImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: Union[torch.Tensor, List[torch.Tensor]]
    """
99
100
    Shape:
    `(batch_size * num_images, 1 + num_patches, num_channels, height, width)`
101

102
103
    Note that `num_patches` may be different per batch and image,
    in which case the data is passed as a list instead of a batched tensor.
104
105
106
107
    """

    image_sizes: torch.Tensor
    """
108
    Shape: `(batch_size * num_images, 2)`
109
110
111
112
113
114
115
116

    This should be in `(height, width)` format.
    """


class Phi3VImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: Union[torch.Tensor, List[torch.Tensor]]
117
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
118
119
120
121
122
123
124
125

    `hidden_size` must match the hidden size of language model backbone.
    """


Phi3VImageInputs = Union[Phi3VImagePixelInputs, Phi3VImageEmbeddingInputs]


126
127
class Phi3ImageEmbeddingBase(nn.Module):

128
    def __init__(self) -> None:
129
130
131
132
133
134
135
136
137
        super().__init__()
        self.layer_idx: int
        self.type_feature: str
        self.img_processor: CLIPVisionModel

    def get_img_features(self,
                         img_embeds: torch.FloatTensor) -> torch.FloatTensor:
        TYPE_FEATURE = self.type_feature

138
139
        # NOTE: we skip the step to select the vision feature layer since
        # this is already done inside the img_processor
140
        img_feature = self.img_processor(img_embeds)
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

        if TYPE_FEATURE == "patch":
            patch_feature = img_feature[:, 1:]
            return patch_feature

        if TYPE_FEATURE == "cls_patch":
            return img_feature

        raise NotImplementedError


# adapted from https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_embedding_phi3_v.py
class Phi3HDImageEmbedding(Phi3ImageEmbeddingBase):
    """Phi3 Image embedding with HD transform."""

156
157
158
159
    def __init__(self,
                 config: PretrainedConfig,
                 quant_config: Optional[QuantizationConfig],
                 prefix: str = "") -> None:
160
        super().__init__()
161
162
163
164
165

        # n_embed or hidden_size
        hidden_size = config.n_embd if hasattr(
            config, 'n_embd') else config.hidden_size

166
167
        self.img_processor = _init_img_processor(
            config, quant_config, prefix=f"{prefix}.img_processor")
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
        image_dim_out = config.img_processor['image_dim_out']
        self.num_img_tokens = config.img_processor['num_img_tokens']

        self.image_dim_out = image_dim_out

        # global_gn and sub_gn for hd transform, serves as line separator
        self.use_hd_transform = config.embd_layer.get('use_hd_transform',
                                                      False)
        self.with_learnable_separator = config.embd_layer.get(
            'with_learnable_separator', False)
        self.hd_transform_order = config.embd_layer.get(
            'hd_transform_order', 'glb_sub')
        # with_hd_transform and with_learnable_separator should have same value
        assert self.use_hd_transform and self.with_learnable_separator

        # 1024 * 4, merge spatial to channel dimension
        self.glb_GN = nn.Parameter(torch.empty([1, 1, self.image_dim_out * 4]))
        self.sub_GN = nn.Parameter(
            torch.empty([1, 1, 1, self.image_dim_out * 4]))

        dim_projection = hidden_size
        depth = 2
        layers = [nn.Linear(image_dim_out * 4, dim_projection)]
        for _ in range(1, depth):
            layers.extend(
                [nn.GELU(),
                 nn.Linear(dim_projection, dim_projection)])
        self.img_projection = nn.Sequential(*layers)

        self.type_feature = config.img_processor.get('type_feature', 'patch')

200
    def forward(self, pixel_values: torch.FloatTensor,
201
                image_sizes: torch.Tensor) -> torch.FloatTensor:
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
        """
        process image and return vision embeddings.

        pixel_values: (num_images, num_crops, c, h, w)
        output: (num_images, num_img_tokens, hidden_size)
        """
        num_images, num_crops, c, h, w = pixel_values.shape
        pixel_values = pixel_values.flatten(0, 1)
        img_features = self.get_img_features(pixel_values)
        img_features = img_features.reshape(num_images, num_crops, -1,
                                            self.image_dim_out)
        image_features_proj = self.hd_feature_transform(
            img_features, image_sizes)
        return image_features_proj

    def hd_feature_transform(self, image_features, image_sizes):
        """
        image_features: (num_images, num_crops+1, 24*24, 1024)
        """
        assert (
            self.hd_transform_order == 'sub_glb'
        ), f'hd_transform_order `{self.hd_transform_order}` not implemented'
        if isinstance(self.img_projection, nn.Sequential):
            target_device = self.img_projection[0].bias.device
            target_dtype = self.img_projection[0].bias.dtype
        else:  # It's a single nn.Linear layer
            target_device = self.img_projection.bias.device
            target_dtype = self.img_projection.bias.dtype

        global_image_features = image_features[:,
                                               0]  # (num_images, 24*24, 1024)
        # global feature can be viewed as a special HD case with num_crops 1x1
        global_image_features_hd = self.reshape_hd_patches_2x2merge(
            global_image_features, 1, 1)
        global_image_features_hd_newline = self.add_image_newline(
            global_image_features_hd)

239
        batch_image_features_proj = []
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
        # need a for loop to process each image because of different image sizes
        # (patch arrangement is different for each image)
        for i, img_size in enumerate(image_sizes):
            h, w = img_size
            h_crop = h // 336
            w_crop = w // 336
            num_crops = h_crop * w_crop

            # NOTE: real num_crops is padded
            # (num_crops, 24*24, 1024)
            sub_image_features = image_features[i, 1:1 + num_crops]
            sub_image_features_hd = self.reshape_hd_patches_2x2merge(
                sub_image_features, h_crop, w_crop)
            sub_image_features_hd_newline = self.add_image_newline(
                sub_image_features_hd)

            # [sub features, separator, global features]
257
258
259
260
261
262
263
264
265
266
267
            image_embeddings = torch.cat([
                sub_image_features_hd_newline.squeeze(
                    0),  # (h_crop*12*(w_crop*12+1), 4096)
                self.glb_GN.squeeze(0),
                global_image_features_hd_newline[i],
            ])
            img_proj = self.img_projection(
                image_embeddings.to(target_device, target_dtype))
            batch_image_features_proj.append(img_proj)

        return batch_image_features_proj
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304

    def reshape_hd_patches_2x2merge(self, image_features, h_crop, w_crop):
        """
        image_features: (num_images*num_crops, 24*24, 1024)
        output: (num_images, h_crop*12, w_crop*12, 4096)
        where h_crop*w_crop == num_crops
        """
        N, L, C = image_features.shape
        assert L == 576 and C == 1024 and N % (h_crop * w_crop) == 0
        num_images = N // (h_crop * w_crop)
        H = int(L**0.5)
        image_features_hd = (
            image_features.reshape(N, H, H, C)  # N, 24, 24, 1024
            .reshape(N, H // 2, 2, H // 2, 2, C)  # N, 12, 2, 12, 2, 1024
            .permute(0, 1, 3, 2, 4, 5)  # N, 12, 12, 2, 2, 1024
            .reshape(N, -1, 4 * C)  # N, 144, 4096
            .reshape(num_images, h_crop, w_crop, H // 2, H // 2,
                     -1)  # n_img, h_crop, w_crop, 12, 12, 4096
            .permute(0, 1, 3, 2, 4, 5)  # n_img, h_crop, 12, w_crop, 12, 4096
            .reshape(num_images, h_crop * H // 2, w_crop * H // 2,
                     4 * C)  # n_img, h_crop*12, w_crop*12, 4096
        )
        return image_features_hd

    def add_image_newline(self, image_features_hd):
        """
        image_features_hd: (num_images, h_crop*12, w_crop*12, 4096)
        output: (num_images, (h_crop*12) * (w_crop*12+1), 4096)
        """
        num_images, h, w, hid_dim = image_features_hd.shape
        # add the newline token to the HD image feature patches
        newline_embeddings = self.sub_GN.expand(num_images, h, -1,
                                                -1)  # (n_img, h, 1, hid_dim)
        image_features_hd_newline = torch.cat(
            [image_features_hd, newline_embeddings],
            dim=2).reshape(num_images, -1, hid_dim)
        return image_features_hd_newline
305
306


307
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L57
308
def _calc_padded_size(*, width: int, height: int, padding_unit: int = 336):
309
310
311
312
313
314
315
316
    target_height = int(np.ceil(height / padding_unit) * padding_unit)
    top_padding = int((target_height - height) / 2)
    bottom_padding = target_height - height - top_padding
    padded_width = width
    padded_height = height + top_padding + bottom_padding
    return padded_width, padded_height


317
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L90
318
def _calc_hd_transform_size(*, width: int, height: int, hd_num: int):
319
320
321
322
323
324
325
326
327
328
329
330
331
332
    transposed = False
    if width < height:
        width, height = height, width
        transposed = True

    ratio = width / height
    scale = 1
    while scale * np.ceil(scale / ratio) <= hd_num:
        scale += 1
    scale -= 1

    new_width = int(scale * 336)
    new_height = int(new_width / ratio)

333
334
    padded_width, padded_height = _calc_padded_size(width=new_width,
                                                    height=new_height)
335
336
337
338
339
340
341

    if transposed:
        padded_width, padded_height = padded_height, padded_width

    return padded_width, padded_height


342
343
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L181
def get_phi3v_image_feature_size(
344
    hf_config: Dict[str, Any],
345
346
347
    *,
    input_height: int,
    input_width: int,
348
    num_crops: int,
349
) -> int:
350
351
    if num_crops is None:
        num_crops = hf_config.get("num_crops", 16)
352
353
354
355
356
357
358
    new_width, new_height = _calc_hd_transform_size(width=input_width,
                                                    height=input_height,
                                                    hd_num=num_crops)

    return (new_height // 336 * new_width // 336 + 1) * 144 + 1 \
        + (new_height // 336 + 1) * 12

359

360
361
362
def get_max_phi3v_image_tokens(ctx: InputContext,
                               *,
                               num_crops: Optional[int] = None):
363
364

    return get_phi3v_image_feature_size(
365
        ctx.get_hf_image_processor_config(),
366
367
        input_height=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
        input_width=MAX_IMAGE_FEATURE_SIZE_WIDTH,
368
        num_crops=num_crops,
369
370
371
    )


372
373
374
375
376
def dummy_data_for_phi3v(ctx: InputContext,
                         seq_len: int,
                         mm_counts: Mapping[str, int],
                         *,
                         num_crops: Optional[int] = None):
377
    num_images = mm_counts["image"]
378

379
    image_feature_size = get_max_phi3v_image_tokens(ctx, num_crops=num_crops)
380

381
    seq_data, ranges = dummy_seq_data_for_clip(
382
383
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
        seq_len,
384
        num_images,
385
        image_token_id=_IMAGE_TOKEN_ID,
386
387
388
389
        image_feature_size_override=image_feature_size,
    )
    mm_data = dummy_image_for_clip(
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
390
        num_images,
391
392
        image_width_override=MAX_IMAGE_FEATURE_SIZE_WIDTH,
        image_height_override=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
393
    )
394

395
    return DummyData(seq_data, mm_data, ranges)
396
397
398


@lru_cache
399
400
401
402
def _get_image_placeholder_token_id_candidates(
    model_config: ModelConfig,
    idx: int,
) -> List[List[int]]:
403
404
405
406
    assert idx > 0

    tokenizer = cached_get_tokenizer(model_config.tokenizer)

407
408
409
410
411
    # This is used when the image token is at the start of the string
    start_candidate = tokenizer.encode(f"<|image_{idx}|>",
                                       add_special_tokens=False)

    # This is used when the image token is in the middle of the string
412
413
414
    # We need to get the token for "<", not "▁<"
    # https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/raw/main/tokenizer.json
    a_token_id, = tokenizer.encode("a", add_special_tokens=False)
415
416
    a_token_id_, *middle_candidate = tokenizer.encode(f"a<|image_{idx}|>",
                                                      add_special_tokens=False)
417
418
    assert a_token_id == a_token_id_

419
    return [start_candidate, middle_candidate]
420
421


422
def input_processor_for_phi3v(ctx: InputContext,
423
                              inputs: DecoderOnlyInputs,
424
425
                              *,
                              num_crops: Optional[int] = None):
426
    multi_modal_data = inputs.get("multi_modal_data")
427
    if multi_modal_data is None or "image" not in multi_modal_data:
428
        return inputs
429

430
    model_config = ctx.model_config
431
    hf_config = ctx.get_hf_image_processor_config()
432
433
434
435

    image_data = multi_modal_data["image"]
    if isinstance(image_data, Image.Image):
        w, h = image_data.size
436
437
438
        image_feature_size = [
            get_phi3v_image_feature_size(hf_config,
                                         input_width=w,
439
440
                                         input_height=h,
                                         num_crops=num_crops)
441
442
443
444
445
446
447
448
449
        ]
        image_data = [image_data]
    elif is_list_of(image_data, Image.Image):
        image_feature_size = []
        for image in image_data:
            w, h = image.size
            image_feature_size.append(
                get_phi3v_image_feature_size(hf_config,
                                             input_width=w,
450
451
                                             input_height=h,
                                             num_crops=num_crops))
452
    elif isinstance(image_data, torch.Tensor):
453
454
        image_feature_size = [image_data.shape[0]]
        image_data = [image_data]
455
    elif is_list_of(image_data, torch.Tensor):
456
        image_feature_size = [item.shape[0] for item in image_data]
457
458
459
    else:
        raise TypeError(f"Invalid image type: {type(image_data)}")

460
    prompt = inputs.get("prompt")
461
    if prompt is None:
462
463
464
        # for async server request, we assume prompt and its token_ids is always
        # in correct format. And num_image_tags == len(image_data) always True.
        image_idx = range(1, len(image_data) + 1)
465
466
        new_prompt = None
    else:
467
        image_idx = sorted(map(int, re.findall(r"<\|image_(\d+)\|>+", prompt)))
468
469
470
471
        if prompt.count("<|image|>") > 0:
            logger.warning("Please follow the prompt format that is "
                           "documented on HuggingFace which does not involve "
                           "repeating <|image|> tokens.")
472
473
474
        elif (num_image_tags := len(image_idx)) > 1:
            assert num_image_tags == len(
                image_data), "The count of image_placeholder not match image's"
475
476
        new_prompt = prompt

477
    prompt_token_ids = inputs["prompt_token_ids"].copy()
478

479
    # masked placeholder with image token id
480
    for idx in image_idx:
481
482
483
484
485
486
487
488
489
490
        candidates = _get_image_placeholder_token_id_candidates(model_config,
                                                                idx=idx)

        for candidate in candidates:
            for i in range(len(prompt_token_ids) - len(candidate) + 1):
                if prompt_token_ids[i:i + len(candidate)] == candidate:
                    prompt_token_ids[i:i +
                                     len(candidate)] = ([_IMAGE_TOKEN_ID] *
                                                        len(candidate))
                    break
491
492
493
494
495
496
497
498
499

    # merge consecutive tag ids
    merged_token_ids: List[int] = []
    for is_placeholder, token_ids in itertools.groupby(
            prompt_token_ids, lambda x: x == _IMAGE_TOKEN_ID):
        if is_placeholder:
            merged_token_ids.append(_IMAGE_TOKEN_ID)
        else:
            merged_token_ids.extend(list(token_ids))
500

501
    # TODO: Move this to utils or integrate with clip.
502
    new_token_ids: List[int] = []
503
504
505
506
507
508
509
510
511
512
    placeholder_idx = 0
    while merged_token_ids:
        token_id = merged_token_ids.pop(0)
        if token_id == _IMAGE_TOKEN_ID:
            new_token_ids.extend(
                repeat_and_pad_token(
                    _IMAGE_TOKEN_ID,
                    repeat_count=image_feature_size[placeholder_idx],
                ))
            placeholder_idx += 1
513
        else:
514
            new_token_ids.append(token_id)
515
516

    # NOTE: Create a defensive copy of the original inputs
517
518
519
    return token_inputs(prompt_token_ids=new_token_ids,
                        prompt=new_prompt,
                        multi_modal_data=multi_modal_data)
520

521
522

@MULTIMODAL_REGISTRY.register_image_input_mapper()
523
@MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_phi3v_image_tokens)
524
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_phi3v)
525
@INPUT_REGISTRY.register_input_processor(input_processor_for_phi3v)
526
class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
527

528
529
530
531
532
    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str = "",
    ) -> None:
533
        super().__init__()
534
535
536
537
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        pooler_config = vllm_config.model_config.pooler_config
        multimodal_config = vllm_config.model_config.multimodal_config
538
        self.config = config
539
        self.multimodal_config = multimodal_config
540
        self.image_token_id = _IMAGE_TOKEN_ID
541

542
543
544
545
546
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            quant_config=quant_config,
547
            prefix="model.embed_tokens",
548
549
550
        )

        # TODO: Optionally initializes this for supporting input embeddings.
551
552
        self.vision_embed_tokens = Phi3HDImageEmbedding(
            config, quant_config, prefix="model.vision_embed_tokens")
553

554
555
        # The prefix is empty intentionally because default prefix of
        # LlamaForCausalLM is "model"
556
557
        self.language_model = LlamaForCausalLM(vllm_config=vllm_config,
                                               prefix="")
558

559
560
        # The same model class supports both language generation and embedding
        # because the architecture name is the same
561
562
563
564
565
        self._pooler = Pooler.from_config_with_defaults(
            pooler_config,
            pooling_type=PoolingType.LAST,
            normalize=True,
            softmax=False)
566
567
568
569
570
571
572
573
        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

    @cached_property
    def sampler(self):
        if hasattr(self.language_model, "sampler"):
            return self.language_model.sampler

Joe Runde's avatar
Joe Runde committed
574
        return get_sampler()
575

576
    def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
577
578
579
580
581
582
583
584
585
586
587
588
589
        expected_dims = (2, )

        def _validate_shape(d: torch.Tensor):
            actual_dims = tuple(d.shape)

            if actual_dims != expected_dims:
                expected_expr = str(expected_dims)
                raise ValueError(
                    f"The expected shape of image sizes per image per batch "
                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")

        for d in data:
            _validate_shape(d)
590
591
592
593
594
595
596

        return data

    def _validate_pixel_values(
        self, data: Union[torch.Tensor, List[torch.Tensor]]
    ) -> Union[torch.Tensor, List[torch.Tensor]]:

597
598
599
600
601
602
603
604
        h = w = CLIP_VIT_LARGE_PATCH14_336_CONFIG.image_size
        expected_dims = (3, h, w)

        def _validate_shape(d: torch.Tensor):
            actual_dims = tuple(d.shape[1:])

            if actual_dims != expected_dims:
                expected_expr = ("num_patches", *map(str, expected_dims))
605
                raise ValueError(
606
                    "The expected shape of pixel values per image per batch "
607
                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
608

609
610
        for d in data:
            _validate_shape(d)
611
612
613

        return data

614
    def _parse_and_validate_image_input(
615
            self, **kwargs: object) -> Optional[Phi3VImageInputs]:
616
617
        pixel_values = kwargs.pop("pixel_values", None)
        image_sizes = kwargs.pop("image_sizes", None)
618
        image_embeds = kwargs.pop("image_embeds", None)
619

620
621
622
623
624
625
626
627
        if pixel_values is None and image_embeds is None:
            return None

        if pixel_values is not None:
            if not isinstance(pixel_values, (torch.Tensor, list)):
                raise ValueError("Incorrect type of pixel values. "
                                 f"Got type: {type(pixel_values)}")

628
            if not isinstance(image_sizes, (torch.Tensor, list)):
629
630
631
632
633
                raise ValueError("Incorrect type of image sizes. "
                                 f"Got type: {type(image_sizes)}")

            return Phi3VImagePixelInputs(
                type="pixel_values",
634
635
636
                data=self._validate_pixel_values(flatten_bn(pixel_values)),
                image_sizes=self._validate_image_sizes(
                    flatten_bn(image_sizes, concat=True)))
637
638
639
640
641

        if image_embeds is not None:
            if not isinstance(image_embeds, torch.Tensor):
                raise ValueError("Incorrect type of image embeddings. "
                                 f"Got type: {type(image_embeds)}")
642

643
644
            return Phi3VImageEmbeddingInputs(
                type="image_embeds",
645
                data=flatten_bn(image_embeds),
646
647
648
649
650
651
652
653
654
655
            )

        raise AssertionError("This line should be unreachable.")

    def _process_image_input(
        self,
        image_input: Phi3VImageInputs,
    ) -> torch.Tensor:

        if image_input["type"] == "image_embeds":
656
657
658
659
660
661
662
663
664
665
666
            image_data = image_input["data"]
            if is_list_of(image_data, torch.Tensor):
                # it's already a list of tensors
                return image_data
            if len(image_data.shape) == 3:
                # 3D tensor
                return list(torch.unbind(image_data, dim=0))
            raise ValueError(
                "We expect batched 2D tensors;"
                "this can be either a list of 2D tensors or a single 3D tensor."
            )
667

668
669
670
        assert self.vision_embed_tokens is not None
        image_embeds = self.vision_embed_tokens(image_input["data"],
                                                image_input["image_sizes"])
671

672
        return image_embeds
673

674
675
676
    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
677
                kv_caches: List[torch.Tensor],
678
679
680
                attn_metadata: AttentionMetadata,
                intermediate_tensors: Optional[IntermediateTensors] = None,
                **kwargs: object):
681
        if intermediate_tensors is not None:
682
            inputs_embeds = None
683
684
685
686
687
        else:
            image_input = self._parse_and_validate_image_input(**kwargs)

            if image_input is not None:
                vision_embeddings = self._process_image_input(image_input)
688
                inputs_embeds = self.embed_tokens(input_ids)
689
690
691
692
                inputs_embeds = merge_multimodal_embeddings(
                    input_ids, inputs_embeds, vision_embeddings,
                    self.image_token_id)
            else:
693
694
695
696
697
698
699
                inputs_embeds = self.language_model.model.embed_tokens(
                    input_ids)

        # always pass the input via `inputs_embeds`
        # to make sure the computation graph is consistent
        # for `torch.compile` integration
        input_ids = None
700

701
702
703
704
705
706
        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  kv_caches,
                                                  attn_metadata,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
707
708
709

        return hidden_states

710
711
712
713
714
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
715
716
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
717
718
719
720
721
722

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
723
        return self.language_model.sample(logits, sampling_metadata)
724

725
726
727
728
729
730
731
    def pooler(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> Optional[PoolerOutput]:
        return self._pooler(hidden_states, pooling_metadata)

732
    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
733
734
        hf_to_vllm_mapper = WeightsMapper(
            orig_to_new_prefix={
735
                "model.vision_embed_tokens.wte": "embed_tokens",
736
737
738
739
740
741
                "model.vision_embed_tokens.": "vision_embed_tokens.",
                "lm_head.": "language_model.lm_head.",
                "model.": "language_model.model.",
            })

        loader = AutoWeightsLoader(self)
742
743
744
745
746
747
748
        autoloaded_weights = loader.load_weights(weights,
                                                 mapper=hf_to_vllm_mapper)

        # The HF config doesn't specify whether these are tied,
        # so we detect it this way
        if "embed_tokens" not in autoloaded_weights:
            self.embed_tokens = self.language_model.model.embed_tokens