phi3v.py 28.2 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
5
6
7
8
9
10
11
12
13
14
15
16
# 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.
pansicheng's avatar
pansicheng committed
17
import re
18
from collections.abc import Iterable, Mapping, Sequence
19
from functools import cached_property
20
from typing import Any, List, Literal, Optional, Set, Tuple, TypedDict, Union
21
22
23

import torch
import torch.nn as nn
24
25
from transformers import (BatchFeature, CLIPVisionConfig, PretrainedConfig,
                          ProcessorMixin)
26

27
from vllm.config import VllmConfig
28
from vllm.logger import init_logger
29
from vllm.model_executor.layers.quantization import QuantizationConfig
Joe Runde's avatar
Joe Runde committed
30
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
31
32
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
33
from vllm.model_executor.sampling_metadata import SamplingMetadata
34
from vllm.multimodal import MULTIMODAL_REGISTRY
35
36
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                    MultiModalKwargs)
37
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
38
                                   ImageSize, MultiModalDataItems)
39
40
# yapf conflicts with isort for this block
# yapf: disable
41
from vllm.multimodal.processing import (BaseMultiModalProcessor,
42
                                        BaseProcessingInfo, BoundPromptUpdate,
43
                                        PlaceholderFeaturesInfo,
44
                                        PromptReplacement, PromptUpdate)
45
# yapf: enable
46
from vllm.multimodal.profiling import BaseDummyInputsBuilder
47
from vllm.sequence import IntermediateTensors
48
from vllm.utils import is_list_of
49

50
from .clip import CLIPVisionModel
51
52
from .interfaces import (MultiModalEmbeddings, SupportsMultiModal, SupportsPP,
                         SupportsQuant)
53
54
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
                    init_vllm_registered_model, maybe_prefix,
55
                    merge_multimodal_embeddings)
56

57
58
logger = init_logger(__name__)

59
60
61
# Cannot find the following 2 numbers from hf config.
_IMAGE_TOKEN_ID = 32044

62
63
64
65
66
67
68
69
70
71
72
73
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)


74
def _init_img_processor(hf_config: PretrainedConfig,
75
76
                        quant_config: Optional[QuantizationConfig],
                        prefix: str = "") -> CLIPVisionModel:
77
78
79
80
81
82
83
84
85
86
87
    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(
88
89
90
        clip_config,
        quant_config,
        num_hidden_layers_override=num_hidden_layers,
91
        prefix=prefix,
92
    )
93
94
95
96

    return img_processor


97
98
99
100
class Phi3VImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: Union[torch.Tensor, List[torch.Tensor]]
    """
101
102
    Shape:
    `(batch_size * num_images, 1 + num_patches, num_channels, height, width)`
103

104
105
    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.
106
107
108
109
    """

    image_sizes: torch.Tensor
    """
110
    Shape: `(batch_size * num_images, 2)`
111
112
113
114
115
116
117
118

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


class Phi3VImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: Union[torch.Tensor, List[torch.Tensor]]
119
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
120
121
122
123
124
125
126
127

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


Phi3VImageInputs = Union[Phi3VImagePixelInputs, Phi3VImageEmbeddingInputs]


128
129
class Phi3ImageEmbeddingBase(nn.Module):

130
    def __init__(self) -> None:
131
132
133
134
135
136
137
138
139
        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

140
141
        # NOTE: we skip the step to select the vision feature layer since
        # this is already done inside the img_processor
142
        img_feature = self.img_processor(img_embeds)
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157

        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."""

158
159
160
161
    def __init__(self,
                 config: PretrainedConfig,
                 quant_config: Optional[QuantizationConfig],
                 prefix: str = "") -> None:
162
        super().__init__()
163
164
165
166
167

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

168
169
        self.img_processor = _init_img_processor(
            config, quant_config, prefix=f"{prefix}.img_processor")
170

171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
        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')

202
    def forward(self, pixel_values: torch.FloatTensor,
203
                image_sizes: torch.Tensor) -> torch.FloatTensor:
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
240
        """
        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)

241
        batch_image_features_proj = []
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
        # 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]
259
260
261
262
263
264
265
266
267
268
269
            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
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
306

    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
307
308


309
class Phi3VProcessingInfo(BaseProcessingInfo):
310

311
    def get_hf_processor(
312
313
314
        self,
        *,
        num_crops: Optional[int] = None,
315
        **kwargs: object,
316
317
    ) -> ProcessorMixin:
        if num_crops is not None:
318
            kwargs["num_crops"] = num_crops
319

320
        return self.ctx.get_hf_processor(**kwargs)
321

322
323
324
    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": None}

325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
        processor: Optional[ProcessorMixin],
    ) -> int:
        if processor is None:
            processor = self.get_hf_processor()

        return processor.calc_num_image_tokens_from_image_size(  # type: ignore
            width=image_width,
            height=image_height,
        )

    def get_image_size_with_most_features(self) -> ImageSize:
341
342
343
        # Result in the max possible feature size (h:w = 16:1)
        return ImageSize(height=8000, width=50)

344
345
346

class Phi3VDummyInputsBuilder(BaseDummyInputsBuilder[Phi3VProcessingInfo]):

347
348
349
350
351
352
353
354
355
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

        hf_processor = self.info.get_hf_processor()
        image_tokens: list[str] = hf_processor.img_tokens  # type: ignore

        return "".join(image_tokens[:num_images])

    def get_dummy_mm_data(
356
        self,
357
358
        seq_len: int,
        mm_counts: Mapping[str, int],
359
    ) -> MultiModalDataDict:
360
        num_images = mm_counts.get("image", 0)
361

362
363
        target_width, target_height = \
            self.info.get_image_size_with_most_features()
364

365
        return {
366
367
368
369
370
371
372
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images)
        }


373
class Phi3VMultiModalProcessor(BaseMultiModalProcessor[Phi3VProcessingInfo]):
374

375
    def _call_hf_processor(
376
377
        self,
        prompt: str,
378
379
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
380
    ) -> BatchFeature:
381
382
        processed_outputs = super()._call_hf_processor(
            prompt=prompt,
383
384
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
385
386
        )

387
388
389
        input_ids = processed_outputs["input_ids"]
        assert isinstance(input_ids, torch.Tensor)

390
391
392
        # Phi3v processor has inserted -1, -2 etc as placeholder in prompt_ids,
        # which will cause OverflowError when decoding the prompt_ids.
        # Therefore, we need to do an early replacement here
393
        input_ids.masked_fill_(input_ids < 0, _IMAGE_TOKEN_ID)
394

395
396
        return processed_outputs

397
398
399
400
401
402
403
404
405
406
407
    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return dict(
            pixel_values=MultiModalFieldConfig.batched("image"),
            image_sizes=MultiModalFieldConfig.batched("image"),
            image_embeds=MultiModalFieldConfig.batched("image"),
        )

408
    def _get_prompt_updates(
409
410
        self,
        mm_items: MultiModalDataItems,
411
        hf_processor_mm_kwargs: Mapping[str, Any],
412
        out_mm_kwargs: MultiModalKwargs,
413
    ) -> Sequence[PromptUpdate]:
414
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
415
416
417
        image_tokens: list[str] = hf_processor.img_tokens  # type: ignore

        def get_replacement_phi3v(item_idx: int):
418
419
420
421
422
423
424
            images = mm_items.get_items(
                "image", (ImageEmbeddingItems, ImageProcessorItems))

            if isinstance(images, ImageEmbeddingItems):
                num_image_tokens = images.get_feature_size(item_idx)
            else:
                image_size = images.get_image_size(item_idx)
425
                num_image_tokens = self.info.get_num_image_tokens(
426
427
                    image_width=image_size.width,
                    image_height=image_size.height,
428
                    processor=hf_processor,
429
430
                )

431
            return [_IMAGE_TOKEN_ID] * num_image_tokens
432

433
434
        num_images = mm_items.get_count("image", strict=False)

435
436
437
438
439
        return [
            PromptReplacement(
                modality="image",
                target=image_token,
                replacement=get_replacement_phi3v,
440
            ) for image_token in image_tokens[:num_images]
441
442
        ]

443
    def _apply_prompt_updates(
444
445
        self,
        token_ids: list[int],
446
        mm_prompt_updates: Mapping[str, Sequence[BoundPromptUpdate]],
447
        mm_item_counts: Mapping[str, int],
448
    ) -> tuple[list[int], str, Mapping[str, list[PlaceholderFeaturesInfo]]]:
pansicheng's avatar
pansicheng committed
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
        # align to hf behavior when there are images
        if len(mm_item_counts):
            tokenizer = self.info.get_tokenizer()
            # to decode token_ids to the original text, we need to
            # 1. remove the first bos token
            # 2. remove space after each special token
            #    introduced by the tokenizer
            if len(token_ids) and token_ids[0] == tokenizer.bos_token_id:
                token_ids = token_ids[1:]
            text = tokenizer.decode(token_ids)
            for special_tokens in tokenizer.special_tokens_map.values():
                if isinstance(special_tokens, str):
                    text = text.replace(f"{special_tokens} ", special_tokens)
                elif isinstance(special_tokens, list):
                    for special_token in special_tokens:
                        text = text.replace(f"{special_token} ", special_token)
            # perform hf behavior
            # https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/64f88b6/processing_phi3_v.py#L407
            pattern = r"<\|image_\d+\|>"
            prompt_chunks = [
                tokenizer(chunk).input_ids
                for chunk in re.split(pattern, text)
            ]
            image_tags = [
                tokenizer(chunk, add_special_tokens=False).input_ids
                for chunk in re.findall(pattern, text)
            ]
            if len(prompt_chunks) > len(image_tags):
                image_tags.append([])
            token_ids = [
                e for sublist in zip(prompt_chunks, image_tags)
                for ele in sublist for e in ele
            ]

483
        token_ids, text, placeholders = super()._apply_prompt_updates(
484
            token_ids=token_ids,
485
            mm_prompt_updates=mm_prompt_updates,
486
487
488
489
490
491
492
            mm_item_counts=mm_item_counts,
        )

        # Keep the behavior in line with HF processor
        if text.startswith("<s> <|image|>"):
            text = text.replace("<s> <|image|>", "<s><|image|>", 1)
            token_ids = [token_ids[0], *token_ids[2:]]
493
494
            placeholders = {
                modality: [
495
                    PlaceholderFeaturesInfo(
496
497
498
                        modality=p.modality,
                        item_idx=p.item_idx,
                        start_idx=p.start_idx - 1,
499
                        tokens=p.tokens,
500
                        is_embed=p.is_embed,
501
502
503
504
                    ) for p in ps
                ]
                for modality, ps in placeholders.items()
            }
505
506
507

        return token_ids, text, placeholders

508

509
510
511
@MULTIMODAL_REGISTRY.register_processor(Phi3VMultiModalProcessor,
                                        info=Phi3VProcessingInfo,
                                        dummy_inputs=Phi3VDummyInputsBuilder)
512
513
class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,
                       SupportsQuant):
514
515
516
517
518
519
520
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            "model.vision_embed_tokens.wte": "embed_tokens",
            "model.vision_embed_tokens.": "vision_embed_tokens.",
            "lm_head.": "language_model.lm_head.",
            "model.": "language_model.model.",
        })
521

522
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
523
        super().__init__()
524
525
        config = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config
526
        self.config = config
527
        self.multimodal_config = multimodal_config
528
        self.image_token_id = _IMAGE_TOKEN_ID
529

530
531
532
533
        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
            org_num_embeddings=config.vocab_size,
534
            quant_config=self.quant_config,
535
            prefix=maybe_prefix(prefix, "model.embed_tokens"),
536
537
538
        )

        # TODO: Optionally initializes this for supporting input embeddings.
539
        self.vision_embed_tokens = Phi3HDImageEmbedding(
540
            config,
541
            self.quant_config,
542
            prefix=maybe_prefix(prefix, "model.vision_embed_tokens"))
543

544
545
546
547
548
549
550
551
552
553
554
        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            # The prefix is empty intentionally because default prefix of
            # LlamaForCausalLM is "model"
            prefix="",
            # We don't directly initialize vLLM's LlamaForCausalLM so we
            # can automatically apply embedding wrapper if this model is
            # initialized as an embedding model
            architectures=["LlamaForCausalLM"],
        )

555
556
557
558
559
560
561
562
        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
563
        return get_sampler()
564

565
    def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
566
567
568
569
570
571
572
573
574
575
576
577
578
        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)
579
580
581
582
583
584
585

        return data

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

586
587
588
589
590
591
592
593
        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))
594
                raise ValueError(
595
                    "The expected shape of pixel values per image per batch "
596
                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
597

598
599
        for d in data:
            _validate_shape(d)
600
601
602

        return data

603
    def _parse_and_validate_image_input(
604
            self, **kwargs: object) -> Optional[Phi3VImageInputs]:
605
606
        pixel_values = kwargs.pop("pixel_values", None)
        image_sizes = kwargs.pop("image_sizes", None)
607
        image_embeds = kwargs.pop("image_embeds", None)
608

609
610
611
612
613
614
615
616
        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)}")

617
            if not isinstance(image_sizes, (torch.Tensor, list)):
618
619
620
621
622
                raise ValueError("Incorrect type of image sizes. "
                                 f"Got type: {type(image_sizes)}")

            return Phi3VImagePixelInputs(
                type="pixel_values",
623
624
625
                data=self._validate_pixel_values(flatten_bn(pixel_values)),
                image_sizes=self._validate_image_sizes(
                    flatten_bn(image_sizes, concat=True)))
626
627
628
629
630

        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)}")
631

632
633
            return Phi3VImageEmbeddingInputs(
                type="image_embeds",
634
                data=flatten_bn(image_embeds),
635
636
637
638
639
640
641
642
643
644
            )

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

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

        if image_input["type"] == "image_embeds":
645
646
647
648
649
650
651
652
            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(
653
                "We expect batched 2D tensors; "
654
655
                "this can be either a list of 2D tensors or a single 3D tensor."
            )
656

657
658
659
        assert self.vision_embed_tokens is not None
        image_embeds = self.vision_embed_tokens(image_input["data"],
                                                image_input["image_sizes"])
660

661
        return image_embeds
662

663
664
665
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

666
    def get_multimodal_embeddings(
667
            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
668
669
670
671
672
673
674
675
676
        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,
677
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
678
679
    ) -> torch.Tensor:
        inputs_embeds = self.embed_tokens(input_ids)
680
        if multimodal_embeddings is not None:
681
            inputs_embeds = merge_multimodal_embeddings(
682
                input_ids, inputs_embeds, multimodal_embeddings,
683
684
685
                self.image_token_id)
        return inputs_embeds

686
687
688
689
    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
                intermediate_tensors: Optional[IntermediateTensors] = None,
690
                inputs_embeds: Optional[torch.Tensor] = None,
691
                **kwargs: object):
692

693
        if intermediate_tensors is not None:
694
            inputs_embeds = None
695
696
697

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility
698
        elif inputs_embeds is None:
699
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
700
701
702
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None
703

704
705
706
707
        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
708
709
710

        return hidden_states

711
712
713
714
715
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
716
717
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
718
719
720
721
722
723

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
724
        return self.language_model.sample(logits, sampling_metadata)
725

726
727
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
728
729

        loader = AutoWeightsLoader(self)
730
        autoloaded_weights = loader.load_weights(weights,
731
                                                 mapper=self.hf_to_vllm_mapper)
732
733
734

        # The HF config doesn't specify whether these are tied,
        # so we detect it this way
735
        if "embed_tokens.weight" not in autoloaded_weights:
736
            self.embed_tokens = self.language_model.model.embed_tokens
737
738
            autoloaded_weights.add("embed_tokens.weight")
        return autoloaded_weights