phi3v.py 29.9 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
from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional, Set,
19
                    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.inputs import NestedTensors, PlaceholderRange
43
from vllm.multimodal.utils import cached_get_tokenizer, repeat_and_pad_token
44
from vllm.sequence import IntermediateTensors, PoolerOutput
45
from vllm.utils import is_list_of
46

47
from .clip import dummy_image_for_clip, dummy_seq_data_for_clip
48
from .interfaces import SupportsMultiModal, SupportsPP
49
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn, maybe_prefix,
50
                    merge_multimodal_embeddings)
51

52
53
logger = init_logger(__name__)

54
55
56
# Cannot find the following 2 numbers from hf config.
_IMAGE_TOKEN_ID = 32044

57
58
59
60
# Result in the max possible feature size (h:w = 16:1)
MAX_IMAGE_FEATURE_SIZE_HEIGHT = 8000
MAX_IMAGE_FEATURE_SIZE_WIDTH = 50

61
62
63
64
65
66
67
68
69
70
71
72
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)


73
def _init_img_processor(hf_config: PretrainedConfig,
74
75
                        quant_config: Optional[QuantizationConfig],
                        prefix: str = "") -> CLIPVisionModel:
76
77
78
79
80
81
82
83
84
85
86
    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(
87
88
89
        clip_config,
        quant_config,
        num_hidden_layers_override=num_hidden_layers,
90
        prefix=prefix,
91
    )
92
93
94
95

    return img_processor


96
97
98
99
class Phi3VImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: Union[torch.Tensor, List[torch.Tensor]]
    """
100
101
    Shape:
    `(batch_size * num_images, 1 + num_patches, num_channels, height, width)`
102

103
104
    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.
105
106
107
108
    """

    image_sizes: torch.Tensor
    """
109
    Shape: `(batch_size * num_images, 2)`
110
111
112
113
114
115
116
117

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


class Phi3VImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: Union[torch.Tensor, List[torch.Tensor]]
118
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
119
120
121
122
123
124
125
126

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


Phi3VImageInputs = Union[Phi3VImagePixelInputs, Phi3VImageEmbeddingInputs]


127
128
class Phi3ImageEmbeddingBase(nn.Module):

129
    def __init__(self) -> None:
130
131
132
133
134
135
136
137
138
        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

139
140
        # NOTE: we skip the step to select the vision feature layer since
        # this is already done inside the img_processor
141
        img_feature = self.img_processor(img_embeds)
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156

        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."""

157
158
159
160
    def __init__(self,
                 config: PretrainedConfig,
                 quant_config: Optional[QuantizationConfig],
                 prefix: str = "") -> None:
161
        super().__init__()
162
163
164
165
166

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

167
168
        self.img_processor = _init_img_processor(
            config, quant_config, prefix=f"{prefix}.img_processor")
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
        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')

201
    def forward(self, pixel_values: torch.FloatTensor,
202
                image_sizes: torch.Tensor) -> torch.FloatTensor:
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
239
        """
        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)

240
        batch_image_features_proj = []
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
        # 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]
258
259
260
261
262
263
264
265
266
267
268
            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
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
305

    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
306
307


308
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L57
309
def _calc_padded_size(*, width: int, height: int, padding_unit: int = 336):
310
311
312
313
314
315
316
317
    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


318
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L90
319
def _calc_hd_transform_size(*, width: int, height: int, hd_num: int):
320
321
322
323
324
325
326
327
328
329
330
331
332
333
    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)

334
335
    padded_width, padded_height = _calc_padded_size(width=new_width,
                                                    height=new_height)
336
337
338
339
340
341
342

    if transposed:
        padded_width, padded_height = padded_height, padded_width

    return padded_width, padded_height


343
344
# 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(
345
    hf_config: Dict[str, Any],
346
347
348
    *,
    input_height: int,
    input_width: int,
349
    num_crops: int,
350
) -> int:
351
352
    if num_crops is None:
        num_crops = hf_config.get("num_crops", 16)
353
354
355
356
357
358
359
    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

360

361
362
363
def get_max_phi3v_image_tokens(ctx: InputContext,
                               *,
                               num_crops: Optional[int] = None):
364
365

    return get_phi3v_image_feature_size(
366
        ctx.get_hf_image_processor_config(),
367
368
        input_height=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
        input_width=MAX_IMAGE_FEATURE_SIZE_WIDTH,
369
        num_crops=num_crops,
370
371
372
    )


373
374
375
376
377
def dummy_data_for_phi3v(ctx: InputContext,
                         seq_len: int,
                         mm_counts: Mapping[str, int],
                         *,
                         num_crops: Optional[int] = None):
378
    num_images = mm_counts["image"]
379

380
    image_feature_size = get_max_phi3v_image_tokens(ctx, num_crops=num_crops)
381

382
    seq_data, ranges = dummy_seq_data_for_clip(
383
384
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
        seq_len,
385
        num_images,
386
        image_token_id=_IMAGE_TOKEN_ID,
387
388
389
390
        image_feature_size_override=image_feature_size,
    )
    mm_data = dummy_image_for_clip(
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
391
        num_images,
392
393
        image_width_override=MAX_IMAGE_FEATURE_SIZE_WIDTH,
        image_height_override=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
394
    )
395

396
    return DummyData(seq_data, mm_data, ranges)
397
398
399


@lru_cache
400
401
402
403
def _get_image_placeholder_token_id_candidates(
    model_config: ModelConfig,
    idx: int,
) -> List[List[int]]:
404
405
406
407
    assert idx > 0

    tokenizer = cached_get_tokenizer(model_config.tokenizer)

408
409
410
411
412
    # 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
413
414
415
    # 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)
416
417
    a_token_id_, *middle_candidate = tokenizer.encode(f"a<|image_{idx}|>",
                                                      add_special_tokens=False)
418
419
    assert a_token_id == a_token_id_

420
    return [start_candidate, middle_candidate]
421
422


423
def input_processor_for_phi3v(ctx: InputContext,
424
                              inputs: DecoderOnlyInputs,
425
426
                              *,
                              num_crops: Optional[int] = None):
427
    multi_modal_data = inputs.get("multi_modal_data")
428
    if multi_modal_data is None or "image" not in multi_modal_data:
429
        return inputs
430

431
    model_config = ctx.model_config
432
    hf_config = ctx.get_hf_image_processor_config()
433
434
435
436

    image_data = multi_modal_data["image"]
    if isinstance(image_data, Image.Image):
        w, h = image_data.size
437
438
439
        image_feature_size = [
            get_phi3v_image_feature_size(hf_config,
                                         input_width=w,
440
441
                                         input_height=h,
                                         num_crops=num_crops)
442
443
444
445
446
447
448
449
450
        ]
        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,
451
452
                                             input_height=h,
                                             num_crops=num_crops))
453
    elif isinstance(image_data, torch.Tensor):
454
455
        image_feature_size = [image_data.shape[0]]
        image_data = [image_data]
456
    elif is_list_of(image_data, torch.Tensor):
457
        image_feature_size = [item.shape[0] for item in image_data]
458
459
460
    else:
        raise TypeError(f"Invalid image type: {type(image_data)}")

461
    prompt = inputs.get("prompt")
462
    if prompt is None:
463
464
465
        # 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)
466
467
        new_prompt = None
    else:
468
        image_idx = sorted(map(int, re.findall(r"<\|image_(\d+)\|>+", prompt)))
469
470
471
472
        if prompt.count("<|image|>") > 0:
            logger.warning("Please follow the prompt format that is "
                           "documented on HuggingFace which does not involve "
                           "repeating <|image|> tokens.")
473
474
475
        elif (num_image_tags := len(image_idx)) > 1:
            assert num_image_tags == len(
                image_data), "The count of image_placeholder not match image's"
476
477
        new_prompt = prompt

478
    prompt_token_ids = inputs["prompt_token_ids"].copy()
479

480
    # masked placeholder with image token id
481
    for idx in image_idx:
482
483
484
485
486
487
488
489
490
491
        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
492
493
494
495
496
497
498
499
500

    # 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))
501

502
    # TODO: Move this to utils or integrate with clip.
503
    new_token_ids: List[int] = []
504
    placeholder_ranges: List[PlaceholderRange] = []
505
506
507
508
    placeholder_idx = 0
    while merged_token_ids:
        token_id = merged_token_ids.pop(0)
        if token_id == _IMAGE_TOKEN_ID:
509
510
511
512
513
514
515
516
517
            replacement_ids = repeat_and_pad_token(
                _IMAGE_TOKEN_ID,
                repeat_count=image_feature_size[placeholder_idx],
            )
            placeholder_ranges.append({
                "offset": len(new_token_ids),
                "length": len(replacement_ids)
            })
            new_token_ids.extend(replacement_ids)
518
            placeholder_idx += 1
519
        else:
520
            new_token_ids.append(token_id)
521
522

    # NOTE: Create a defensive copy of the original inputs
523
524
    return token_inputs(prompt_token_ids=new_token_ids,
                        prompt=new_prompt,
525
526
                        multi_modal_data=multi_modal_data,
                        multi_modal_placeholders={"image": placeholder_ranges})
527

528
529

@MULTIMODAL_REGISTRY.register_image_input_mapper()
530
@MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_phi3v_image_tokens)
531
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_phi3v)
532
@INPUT_REGISTRY.register_input_processor(input_processor_for_phi3v)
533
class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
534

535
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
536
        super().__init__()
537
538
539
540
        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
541
        self.config = config
542
        self.multimodal_config = multimodal_config
543
        self.image_token_id = _IMAGE_TOKEN_ID
544

545
546
547
548
549
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
            quant_config=quant_config,
550
            prefix=maybe_prefix(prefix, "model.embed_tokens"),
551
552
553
        )

        # TODO: Optionally initializes this for supporting input embeddings.
554
        self.vision_embed_tokens = Phi3HDImageEmbedding(
555
556
557
            config,
            quant_config,
            prefix=maybe_prefix(prefix, "model.vision_embed_tokens"))
558

559
560
        # The prefix is empty intentionally because default prefix of
        # LlamaForCausalLM is "model"
561
562
        self.language_model = LlamaForCausalLM(vllm_config=vllm_config,
                                               prefix="")
563

564
565
        # The same model class supports both language generation and embedding
        # because the architecture name is the same
566
567
568
569
570
        self._pooler = Pooler.from_config_with_defaults(
            pooler_config,
            pooling_type=PoolingType.LAST,
            normalize=True,
            softmax=False)
571
572
573
574
575
576
577
578
        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
579
        return get_sampler()
580

581
    def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
582
583
584
585
586
587
588
589
590
591
592
593
594
        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)
595
596
597
598
599
600
601

        return data

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

602
603
604
605
606
607
608
609
        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))
610
                raise ValueError(
611
                    "The expected shape of pixel values per image per batch "
612
                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
613

614
615
        for d in data:
            _validate_shape(d)
616
617
618

        return data

619
    def _parse_and_validate_image_input(
620
            self, **kwargs: object) -> Optional[Phi3VImageInputs]:
621
622
        pixel_values = kwargs.pop("pixel_values", None)
        image_sizes = kwargs.pop("image_sizes", None)
623
        image_embeds = kwargs.pop("image_embeds", None)
624

625
626
627
628
629
630
631
632
        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)}")

633
            if not isinstance(image_sizes, (torch.Tensor, list)):
634
635
636
637
638
                raise ValueError("Incorrect type of image sizes. "
                                 f"Got type: {type(image_sizes)}")

            return Phi3VImagePixelInputs(
                type="pixel_values",
639
640
641
                data=self._validate_pixel_values(flatten_bn(pixel_values)),
                image_sizes=self._validate_image_sizes(
                    flatten_bn(image_sizes, concat=True)))
642
643
644
645
646

        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)}")
647

648
649
            return Phi3VImageEmbeddingInputs(
                type="image_embeds",
650
                data=flatten_bn(image_embeds),
651
652
653
654
655
656
657
658
659
660
            )

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

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

        if image_input["type"] == "image_embeds":
661
662
663
664
665
666
667
668
669
670
671
            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."
            )
672

673
674
675
        assert self.vision_embed_tokens is not None
        image_embeds = self.vision_embed_tokens(image_input["data"],
                                                image_input["image_sizes"])
676

677
        return image_embeds
678

679
    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
680
681
682
683
684
685
686
687
688
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return None
        vision_embeddings = self._process_image_input(image_input)
        return vision_embeddings

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
689
        multimodal_embeddings: Optional[NestedTensors] = None,
690
691
    ) -> torch.Tensor:
        inputs_embeds = self.embed_tokens(input_ids)
692
        if multimodal_embeddings is not None:
693
            inputs_embeds = merge_multimodal_embeddings(
694
                input_ids, inputs_embeds, multimodal_embeddings,
695
696
697
                self.image_token_id)
        return inputs_embeds

698
699
700
    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
701
                kv_caches: List[torch.Tensor],
702
703
                attn_metadata: AttentionMetadata,
                intermediate_tensors: Optional[IntermediateTensors] = None,
704
                inputs_embeds: Optional[torch.Tensor] = None,
705
                **kwargs: object):
706

707
        if intermediate_tensors is not None:
708
            inputs_embeds = None
709
710
711

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility
712
        elif inputs_embeds is None:
713
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
714
715
716
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None
717

718
719
720
721
722
723
        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  kv_caches,
                                                  attn_metadata,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
724
725
726

        return hidden_states

727
728
729
730
731
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
732
733
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
734
735
736
737
738
739

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
740
        return self.language_model.sample(logits, sampling_metadata)
741

742
743
744
745
746
747
748
    def pooler(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> Optional[PoolerOutput]:
        return self._pooler(hidden_states, pooling_metadata)

749
750
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
751
752
        hf_to_vllm_mapper = WeightsMapper(
            orig_to_new_prefix={
753
                "model.vision_embed_tokens.wte": "embed_tokens",
754
755
756
757
758
759
                "model.vision_embed_tokens.": "vision_embed_tokens.",
                "lm_head.": "language_model.lm_head.",
                "model.": "language_model.model.",
            })

        loader = AutoWeightsLoader(self)
760
761
762
763
764
        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
765
        if "embed_tokens.weight" not in autoloaded_weights:
766
            self.embed_tokens = self.language_model.model.embed_tokens
767
768
            autoloaded_weights.add("embed_tokens.weight")
        return autoloaded_weights