phi3v.py 29.4 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.quantization import QuantizationConfig
Joe Runde's avatar
Joe Runde committed
33
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
34
35
from vllm.model_executor.layers.vocab_parallel_embedding import (
    VocabParallelEmbedding)
36
from vllm.model_executor.models.clip import CLIPVisionModel
37
from vllm.model_executor.sampling_metadata import SamplingMetadata
38
from vllm.multimodal import MULTIMODAL_REGISTRY
39
from vllm.multimodal.inputs import NestedTensors, PlaceholderRange
40
from vllm.multimodal.utils import cached_get_tokenizer, repeat_and_pad_token
41
from vllm.sequence import IntermediateTensors
42
from vllm.utils import is_list_of
43

44
from .clip import dummy_image_for_clip, dummy_seq_data_for_clip
45
from .interfaces import SupportsMultiModal, SupportsPP
46
47
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
                    init_vllm_registered_model, maybe_prefix,
48
                    merge_multimodal_embeddings)
49

50
51
logger = init_logger(__name__)

52
53
54
# Cannot find the following 2 numbers from hf config.
_IMAGE_TOKEN_ID = 32044

55
56
57
58
# Result in the max possible feature size (h:w = 16:1)
MAX_IMAGE_FEATURE_SIZE_HEIGHT = 8000
MAX_IMAGE_FEATURE_SIZE_WIDTH = 50

59
60
61
62
63
64
65
66
67
68
69
70
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)


71
def _init_img_processor(hf_config: PretrainedConfig,
72
73
                        quant_config: Optional[QuantizationConfig],
                        prefix: str = "") -> CLIPVisionModel:
74
75
76
77
78
79
80
81
82
83
84
    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(
85
86
87
        clip_config,
        quant_config,
        num_hidden_layers_override=num_hidden_layers,
88
        prefix=prefix,
89
    )
90
91
92
93

    return img_processor


94
95
96
97
class Phi3VImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: Union[torch.Tensor, List[torch.Tensor]]
    """
98
99
    Shape:
    `(batch_size * num_images, 1 + num_patches, num_channels, height, width)`
100

101
102
    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.
103
104
105
106
    """

    image_sizes: torch.Tensor
    """
107
    Shape: `(batch_size * num_images, 2)`
108
109
110
111
112
113
114
115

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


class Phi3VImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: Union[torch.Tensor, List[torch.Tensor]]
116
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
117
118
119
120
121
122
123
124

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


Phi3VImageInputs = Union[Phi3VImagePixelInputs, Phi3VImageEmbeddingInputs]


125
126
class Phi3ImageEmbeddingBase(nn.Module):

127
    def __init__(self) -> None:
128
129
130
131
132
133
134
135
136
        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

137
138
        # NOTE: we skip the step to select the vision feature layer since
        # this is already done inside the img_processor
139
        img_feature = self.img_processor(img_embeds)
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154

        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."""

155
156
157
158
    def __init__(self,
                 config: PretrainedConfig,
                 quant_config: Optional[QuantizationConfig],
                 prefix: str = "") -> None:
159
        super().__init__()
160
161
162
163
164

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

165
166
        self.img_processor = _init_img_processor(
            config, quant_config, prefix=f"{prefix}.img_processor")
167

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
        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')

199
    def forward(self, pixel_values: torch.FloatTensor,
200
                image_sizes: torch.Tensor) -> torch.FloatTensor:
201
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
        """
        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)

238
        batch_image_features_proj = []
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
        # 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]
256
257
258
259
260
261
262
263
264
265
266
            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
267
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

    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
304
305


306
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L57
307
def _calc_padded_size(*, width: int, height: int, padding_unit: int = 336):
308
309
310
311
312
313
314
315
    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


316
# Based on https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/blob/main/image_processing_phi3_v.py#L90
317
def _calc_hd_transform_size(*, width: int, height: int, hd_num: int):
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    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)

332
333
    padded_width, padded_height = _calc_padded_size(width=new_width,
                                                    height=new_height)
334
335
336
337
338
339
340

    if transposed:
        padded_width, padded_height = padded_height, padded_width

    return padded_width, padded_height


341
342
# 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(
343
    hf_config: Dict[str, Any],
344
345
346
    *,
    input_height: int,
    input_width: int,
347
    num_crops: int,
348
) -> int:
349
350
    if num_crops is None:
        num_crops = hf_config.get("num_crops", 16)
351
352
353
354
355
356
357
    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

358

359
360
361
def get_max_phi3v_image_tokens(ctx: InputContext,
                               *,
                               num_crops: Optional[int] = None):
362
363

    return get_phi3v_image_feature_size(
364
        ctx.get_hf_image_processor_config(),
365
366
        input_height=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
        input_width=MAX_IMAGE_FEATURE_SIZE_WIDTH,
367
        num_crops=num_crops,
368
369
370
    )


371
372
373
374
375
def dummy_data_for_phi3v(ctx: InputContext,
                         seq_len: int,
                         mm_counts: Mapping[str, int],
                         *,
                         num_crops: Optional[int] = None):
376
    num_images = mm_counts["image"]
377

378
    image_feature_size = get_max_phi3v_image_tokens(ctx, num_crops=num_crops)
379

380
    seq_data, ranges = dummy_seq_data_for_clip(
381
382
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
        seq_len,
383
        num_images,
384
        image_token_id=_IMAGE_TOKEN_ID,
385
386
387
388
        image_feature_size_override=image_feature_size,
    )
    mm_data = dummy_image_for_clip(
        CLIP_VIT_LARGE_PATCH14_336_CONFIG,
389
        num_images,
390
391
        image_width_override=MAX_IMAGE_FEATURE_SIZE_WIDTH,
        image_height_override=MAX_IMAGE_FEATURE_SIZE_HEIGHT,
392
    )
393

394
    return DummyData(seq_data, mm_data, ranges)
395
396
397


@lru_cache
398
399
400
401
def _get_image_placeholder_token_id_candidates(
    model_config: ModelConfig,
    idx: int,
) -> List[List[int]]:
402
403
404
405
    assert idx > 0

    tokenizer = cached_get_tokenizer(model_config.tokenizer)

406
407
408
409
410
    # 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
411
412
413
    # 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)
414
415
    a_token_id_, *middle_candidate = tokenizer.encode(f"a<|image_{idx}|>",
                                                      add_special_tokens=False)
416
417
    assert a_token_id == a_token_id_

418
    return [start_candidate, middle_candidate]
419
420


421
def input_processor_for_phi3v(ctx: InputContext,
422
                              inputs: DecoderOnlyInputs,
423
424
                              *,
                              num_crops: Optional[int] = None):
425
    multi_modal_data = inputs.get("multi_modal_data")
426
    if multi_modal_data is None or "image" not in multi_modal_data:
427
        return inputs
428

429
    model_config = ctx.model_config
430
    hf_config = ctx.get_hf_image_processor_config()
431
432
433
434

    image_data = multi_modal_data["image"]
    if isinstance(image_data, Image.Image):
        w, h = image_data.size
435
436
437
        image_feature_size = [
            get_phi3v_image_feature_size(hf_config,
                                         input_width=w,
438
439
                                         input_height=h,
                                         num_crops=num_crops)
440
441
442
443
444
445
446
447
448
        ]
        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,
449
450
                                             input_height=h,
                                             num_crops=num_crops))
451
    elif isinstance(image_data, torch.Tensor):
452
453
        image_feature_size = [image_data.shape[0]]
        image_data = [image_data]
454
    elif is_list_of(image_data, torch.Tensor):
455
        image_feature_size = [item.shape[0] for item in image_data]
456
457
458
    else:
        raise TypeError(f"Invalid image type: {type(image_data)}")

459
    prompt = inputs.get("prompt")
460
    if prompt is None:
461
462
463
        # 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)
464
465
        new_prompt = None
    else:
466
        image_idx = sorted(map(int, re.findall(r"<\|image_(\d+)\|>+", prompt)))
467
468
469
470
        if prompt.count("<|image|>") > 0:
            logger.warning("Please follow the prompt format that is "
                           "documented on HuggingFace which does not involve "
                           "repeating <|image|> tokens.")
471
472
473
        elif (num_image_tags := len(image_idx)) > 1:
            assert num_image_tags == len(
                image_data), "The count of image_placeholder not match image's"
474
475
        new_prompt = prompt

476
    prompt_token_ids = inputs["prompt_token_ids"].copy()
477

478
    # masked placeholder with image token id
479
    for idx in image_idx:
480
481
482
483
484
485
486
487
488
489
        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
490
491
492
493
494
495
496
497
498

    # 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))
499

500
    # TODO: Move this to utils or integrate with clip.
501
    new_token_ids: List[int] = []
502
    placeholder_ranges: List[PlaceholderRange] = []
503
504
505
506
    placeholder_idx = 0
    while merged_token_ids:
        token_id = merged_token_ids.pop(0)
        if token_id == _IMAGE_TOKEN_ID:
507
508
509
510
511
512
513
514
515
            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)
516
            placeholder_idx += 1
517
        else:
518
            new_token_ids.append(token_id)
519
520

    # NOTE: Create a defensive copy of the original inputs
521
522
    return token_inputs(prompt_token_ids=new_token_ids,
                        prompt=new_prompt,
523
524
                        multi_modal_data=multi_modal_data,
                        multi_modal_placeholders={"image": placeholder_ranges})
525

526
527

@MULTIMODAL_REGISTRY.register_image_input_mapper()
528
@MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_phi3v_image_tokens)
529
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_phi3v)
530
@INPUT_REGISTRY.register_input_processor(input_processor_for_phi3v)
531
class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
532

533
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
534
        super().__init__()
535
536
537
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_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=maybe_prefix(prefix, "model.embed_tokens"),
548
549
550
        )

        # TODO: Optionally initializes this for supporting input embeddings.
551
        self.vision_embed_tokens = Phi3HDImageEmbedding(
552
553
554
            config,
            quant_config,
            prefix=maybe_prefix(prefix, "model.vision_embed_tokens"))
555

556
557
558
559
560
561
562
563
564
565
566
        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"],
        )

567
568
569
570
571
572
573
574
        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
575
        return get_sampler()
576

577
    def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
578
579
580
581
582
583
584
585
586
587
588
589
590
        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)
591
592
593
594
595
596
597

        return data

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

598
599
600
601
602
603
604
605
        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))
606
                raise ValueError(
607
                    "The expected shape of pixel values per image per batch "
608
                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
609

610
611
        for d in data:
            _validate_shape(d)
612
613
614

        return data

615
    def _parse_and_validate_image_input(
616
            self, **kwargs: object) -> Optional[Phi3VImageInputs]:
617
618
        pixel_values = kwargs.pop("pixel_values", None)
        image_sizes = kwargs.pop("image_sizes", None)
619
        image_embeds = kwargs.pop("image_embeds", None)
620

621
622
623
624
625
626
627
628
        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)}")

629
            if not isinstance(image_sizes, (torch.Tensor, list)):
630
631
632
633
634
                raise ValueError("Incorrect type of image sizes. "
                                 f"Got type: {type(image_sizes)}")

            return Phi3VImagePixelInputs(
                type="pixel_values",
635
636
637
                data=self._validate_pixel_values(flatten_bn(pixel_values)),
                image_sizes=self._validate_image_sizes(
                    flatten_bn(image_sizes, concat=True)))
638
639
640
641
642

        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)}")
643

644
645
            return Phi3VImageEmbeddingInputs(
                type="image_embeds",
646
                data=flatten_bn(image_embeds),
647
648
649
650
651
652
653
654
655
656
            )

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

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

        if image_input["type"] == "image_embeds":
657
658
659
660
661
662
663
664
665
666
667
            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."
            )
668

669
670
671
        assert self.vision_embed_tokens is not None
        image_embeds = self.vision_embed_tokens(image_input["data"],
                                                image_input["image_sizes"])
672

673
        return image_embeds
674

675
    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
676
677
678
679
680
681
682
683
684
        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,
685
        multimodal_embeddings: Optional[NestedTensors] = None,
686
687
    ) -> torch.Tensor:
        inputs_embeds = self.embed_tokens(input_ids)
688
        if multimodal_embeddings is not None:
689
            inputs_embeds = merge_multimodal_embeddings(
690
                input_ids, inputs_embeds, multimodal_embeddings,
691
692
693
                self.image_token_id)
        return inputs_embeds

694
695
696
    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
697
                kv_caches: List[torch.Tensor],
698
699
                attn_metadata: AttentionMetadata,
                intermediate_tensors: Optional[IntermediateTensors] = None,
700
                inputs_embeds: Optional[torch.Tensor] = None,
701
                **kwargs: object):
702

703
        if intermediate_tensors is not None:
704
            inputs_embeds = None
705
706
707

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility
708
        elif inputs_embeds is None:
709
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
710
711
712
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None
713

714
715
716
717
718
719
        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  kv_caches,
                                                  attn_metadata,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)
720
721
722

        return hidden_states

723
724
725
726
727
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
728
729
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
730
731
732
733
734
735

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
736
        return self.language_model.sample(logits, sampling_metadata)
737

738
739
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
740
741
        hf_to_vllm_mapper = WeightsMapper(
            orig_to_new_prefix={
742
                "model.vision_embed_tokens.wte": "embed_tokens",
743
744
745
746
747
748
                "model.vision_embed_tokens.": "vision_embed_tokens.",
                "lm_head.": "language_model.lm_head.",
                "model.": "language_model.model.",
            })

        loader = AutoWeightsLoader(self)
749
750
751
752
753
        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
754
        if "embed_tokens.weight" not in autoloaded_weights:
755
            self.embed_tokens = self.language_model.model.embed_tokens
756
757
            autoloaded_weights.add("embed_tokens.weight")
        return autoloaded_weights