minicpmv.py 52.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.
Alphi's avatar
Alphi committed
25
"""Inference-only MiniCPM-V model compatible with HuggingFace weights."""
26
import math
27
from collections import defaultdict
28
from collections.abc import Iterable, Mapping, Sequence
29
from functools import partial
30
from typing import Annotated, Any, Callable, Literal, Optional, Union
31

32
import numpy as np
33
import torch
Alphi's avatar
Alphi committed
34
import torch.types
35
from torch import nn
36
from transformers import BatchFeature, PretrainedConfig
37
from typing_extensions import TypeVar
38

39
from vllm.config import VllmConfig
40
from vllm.model_executor.layers.quantization import QuantizationConfig
tc-mb's avatar
tc-mb committed
41
42
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
43
from vllm.model_executor.layers.resampler import (BaseResampler, Resampler2,
44
                                                  get_2d_sincos_pos_embed)
Jee Jee Li's avatar
Jee Jee Li committed
45
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
46
47
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.model_executor.models.minicpm import MiniCPMForCausalLM
48
from vllm.model_executor.models.module_mapping import MultiModelKeys
49
from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM
50
from vllm.model_executor.sampling_metadata import SamplingMetadata
51
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
52
53
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                    NestedTensors)
54
55
from vllm.multimodal.parse import (DictEmbeddingItems, ImageItem,
                                   ImageProcessorItems, ImageSize,
56
57
                                   ModalityData, ModalityDataItems,
                                   MultiModalDataItems, MultiModalDataParser,
58
                                   VideoItem, VideoProcessorItems)
59
from vllm.multimodal.processing import (BaseMultiModalProcessor,
60
                                        BaseProcessingInfo, PromptReplacement,
61
                                        PromptUpdate, PromptUpdateDetails)
62
from vllm.multimodal.profiling import BaseDummyInputsBuilder
63
from vllm.platforms import current_platform
64
from vllm.sequence import IntermediateTensors
65
from vllm.utils import flatten_2d_lists
66
from vllm.utils.tensor_schema import TensorSchema, TensorShape
67

Jee Jee Li's avatar
Jee Jee Li committed
68
from .idefics2_vision_model import Idefics2VisionTransformer
69
70
71
72
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
                         SupportsMultiModal, SupportsPP)
from .utils import (AutoWeightsLoader, flatten_bn, maybe_prefix,
                    merge_multimodal_embeddings)
73

74
75
76
# For profile run
_MAX_FRAMES_PER_VIDEO = 16

77

78
class MiniCPMVImagePixelInputs(TensorSchema):
Jee Jee Li's avatar
Jee Jee Li committed
79
    """
80
81
82
83
84
85
    Dimensions:
        - bns: Batch size * number of images * number of slices
        - bn: Batch size * number of images
        - c: Number of channels
        - h: Height
        - w: Width
Jee Jee Li's avatar
Jee Jee Li committed
86
87
    """

88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
    def _validate_nested_tensors(
        self,
        value: Union[list[torch.Tensor], tuple[torch.Tensor, ...]],
        field_name: str,
        expected_shape: tuple[Union[int, str], ...],
        dynamic_dims: set[str],
    ) -> tuple[int, ...]:
        # value[0] is the scaled image,
        # and value[1:] is a collection of image slices.
        # It is ensured that all slices in the collection
        # have the same shape.
        if field_name == "pixel_values":
            value = value[1:] if len(value) > 1 else value

        return super()._validate_nested_tensors(value, field_name,
                                                expected_shape, dynamic_dims)

105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
    type: Literal["pixel_values"] = "pixel_values"

    # Note that the image size may vary, so we pass it as a list instead of a
    # batched tensor.
    pixel_values: Annotated[
        list[torch.Tensor],
        TensorShape("bns", "c", "h", "w"),
    ]
    tgt_sizes: Annotated[
        torch.Tensor,
        TensorShape("bns", 2),  # This should be in `(height, width)` format.
    ]
    num_slices: Annotated[
        torch.Tensor,
        TensorShape("bn"),
    ]


class MiniCPMVImageEmbeddingInputs(TensorSchema):
Jee Jee Li's avatar
Jee Jee Li committed
124
    """
125
126
127
128
    Dimensions:
        - bn: Batch size * number of images
        - ns: Number of slices
        - hs: Hidden size (must match language model backbone)
Jee Jee Li's avatar
Jee Jee Li committed
129
130
    """

131
    type: Literal["image_embeds"]
132
133
134
135
    image_embeds: Annotated[
        Union[torch.Tensor, list[torch.Tensor]],
        TensorShape("bn", "ns", "hs"),
    ]
136
137
138
139
140


MiniCPMVImageInputs = Union[MiniCPMVImagePixelInputs,
                            MiniCPMVImageEmbeddingInputs]

Jee Jee Li's avatar
Jee Jee Li committed
141
142
143
144
145
DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)


class Resampler2_5(BaseResampler):

146
147
148
149
150
151
    def __init__(self,
                 num_queries: int,
                 embed_dim: int,
                 num_heads: int,
                 kv_dim: Optional[int] = None,
                 norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
152
                 max_size: tuple[int, int] = (70, 70),
153
154
155
156
157
158
159
160
161
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = "") -> None:
        super().__init__(num_queries,
                         embed_dim,
                         num_heads,
                         kv_dim,
                         norm_layer,
                         quant_config=quant_config,
                         prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
162
163
164

        self.max_size = max_size
        self._set_2d_pos_cache(self.max_size)
165

Alphi's avatar
Alphi committed
166
    def _set_2d_pos_cache(self,
167
                          max_size: tuple[int, int],
Jee Jee Li's avatar
Jee Jee Li committed
168
169
170
171
172
                          device: torch.types.Device = "cpu") -> None:
        pos_embed_arr = get_2d_sincos_pos_embed(self.embed_dim,
                                                max_size,
                                                version=(2, 5))
        pos_embed = torch.from_numpy(pos_embed_arr).float().to(device)
173
174
        self.register_buffer("pos_embed", pos_embed, persistent=False)

Alphi's avatar
Alphi committed
175
    def _adjust_pos_cache(self, tgt_sizes: torch.Tensor,
Jee Jee Li's avatar
Jee Jee Li committed
176
177
178
179
180
                          device: torch.types.Device) -> None:
        max_h = tgt_sizes[:, 0].max().item()
        max_w = tgt_sizes[:, 1].max().item()
        assert isinstance(max_h, int) and isinstance(max_w, int)

181
        if max_h > self.max_size[0] or max_w > self.max_size[1]:
Jee Jee Li's avatar
Jee Jee Li committed
182
            self.max_size = (
183
                max(max_h, self.max_size[0]),
Jee Jee Li's avatar
Jee Jee Li committed
184
185
                max(max_w, self.max_size[1]),
            )
186
187
            self._set_2d_pos_cache(self.max_size, device)

Jee Jee Li's avatar
Jee Jee Li committed
188
189
    def forward(self, x: torch.Tensor,
                tgt_sizes: torch.Tensor) -> torch.Tensor:
190
191
192
193
194
195
196
197
198
199
        assert x.shape[0] == tgt_sizes.shape[0]
        bs = x.shape[0]

        device = x.device
        dtype = x.dtype

        patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]

        self._adjust_pos_cache(tgt_sizes, device=device)

Jee Jee Li's avatar
Jee Jee Li committed
200
201
202
        max_patch_len = patch_len.max().item()
        assert isinstance(max_patch_len, int)

203
204
205
206
207
208
        key_padding_mask = torch.zeros((bs, max_patch_len),
                                       dtype=torch.bool,
                                       device=device)

        pos_embed = []
        for i in range(bs):
Jee Jee Li's avatar
Jee Jee Li committed
209
            tgt_h, tgt_w = tgt_sizes[i].tolist()
210
211
212
213
214
215
216
217
            pos_embed.append(self.pos_embed[:tgt_h, :tgt_w, :].reshape(
                (tgt_h * tgt_w, -1)).to(dtype))  # patches * D
            key_padding_mask[i, patch_len[i]:] = True
        pos_embed = torch.nn.utils.rnn.pad_sequence(pos_embed,
                                                    batch_first=True,
                                                    padding_value=0.0).permute(
                                                        1, 0,
                                                        2)  # BLD => L * B * D
Jee Jee Li's avatar
Jee Jee Li committed
218
        x, _ = self.kv_proj(x)  # B * L * D
219
220
221
222
223
224
225
226
        x = self.ln_kv(x).permute(1, 0, 2)  # L * B * D

        q = self.ln_q(self.query)  # Q * D

        out = self.attn(
            self._repeat(q, bs),  # Q * B * D
            x + pos_embed,  # L * B * D +  L * B * D
            x,
Jee Jee Li's avatar
Jee Jee Li committed
227
228
            key_padding_mask=key_padding_mask,
        )[0]
229
230
231
232
233
234
235
236
        #  out: Q * B * D
        x = out.permute(1, 0, 2)  # B * Q * D

        x = self.ln_post(x)
        x = x @ self.proj
        return x


237
def get_version_by_config(config: PretrainedConfig) -> tuple[int, ...]:
238
239
240
241
242
243
244
245
246
247
248
249
    version_float = getattr(config, "version", None)

    # The old configs do not include version number
    # TODO: Remove this after the HF repos are updated
    if version_float is None:
        if config.hidden_size == 2304 and config.query_num == 64:
            return (2, 0)
        return (2, 5)
    version_str = str(version_float)
    return tuple(int(x) for x in version_str.split("."))


250
def _minicpmv_field_config(hf_inputs: Mapping[str, torch.Tensor]):
251
252
253
254
255
256
    pixel_values = hf_inputs.get("pixel_values", torch.empty(0))
    num_images = len(pixel_values)

    video_pixel_values = hf_inputs.get("video_pixel_values", torch.empty(0))
    num_videos = len(video_pixel_values)

257
    return dict(
258
        pixel_values=MultiModalFieldConfig.batched("image"),
259
        image_sizes=MultiModalFieldConfig.batched("image"),
260
261
262
        tgt_sizes=MultiModalFieldConfig.batched("image"),
        image_embeds=MultiModalFieldConfig.batched("image"),
        video_pixel_values=MultiModalFieldConfig.batched("video"),
263
        video_image_sizes=MultiModalFieldConfig.batched("video"),
264
265
        video_tgt_sizes=MultiModalFieldConfig.batched("video"),
        video_embeds=MultiModalFieldConfig.batched("video"),
266
267
        image_token_id=MultiModalFieldConfig.shared("image", num_images),
        video_token_id=MultiModalFieldConfig.shared("video", num_videos),
268
269
270
271
272
273
274
275
    )


class MiniCPMVImageEmbeddingItems(DictEmbeddingItems):

    def __init__(
        self,
        data: Mapping[str, torch.Tensor],
276
277
278
279
        fields_factory: Callable[
            [Mapping[str, torch.Tensor]],
            Mapping[str, MultiModalFieldConfig],
        ],
280
281
282
283
284
    ) -> None:
        super().__init__(
            data,
            modality="image",
            required_fields={"image_embeds", "image_sizes"},
285
            fields_factory=fields_factory,
286
287
288
289
290
291
292
293
294
295
296
297
        )

    def get_image_size(self, index: int) -> ImageSize:
        image_size = self.get(index)["image_sizes"].tolist()
        return ImageSize(width=image_size[0], height=image_size[1])


class MiniCPMVVideoEmbeddingItems(DictEmbeddingItems):

    def __init__(
        self,
        data: Mapping[str, torch.Tensor],
298
299
300
301
        fields_factory: Callable[
            [Mapping[str, torch.Tensor]],
            Mapping[str, MultiModalFieldConfig],
        ],
302
303
304
305
306
    ) -> None:
        super().__init__(
            data,
            modality="video",
            required_fields={"video_embeds", "video_image_sizes"},
307
            fields_factory=fields_factory,
308
309
310
311
312
313
314
315
316
317
        )

    def get_frame_size(self, index: int) -> ImageSize:
        frame_size = self.get(index)["video_image_sizes"].tolist()
        return ImageSize(width=frame_size[0], height=frame_size[1])

    def get_num_frames(self, index: int) -> int:
        return len(self.get(index)["video_image_sizes"])


318
319
320
321
322
class MiniCPMVMultiModalDataParser(MultiModalDataParser):

    def _parse_image_data(
        self,
        data: Union[dict[str, torch.Tensor], ModalityData[ImageItem]],
323
    ) -> Optional[ModalityDataItems[Any, Any]]:
324
        if isinstance(data, dict):
325
326
            return MiniCPMVImageEmbeddingItems(
                data,
327
                fields_factory=_minicpmv_field_config,
328
329
            )

330
331
332
333
334
        return super()._parse_image_data(data)

    def _parse_video_data(
        self,
        data: Union[dict[str, torch.Tensor], ModalityData[VideoItem]],
335
    ) -> Optional[ModalityDataItems[Any, Any]]:
336
        if isinstance(data, dict):
337
338
            return MiniCPMVVideoEmbeddingItems(
                data,
339
                fields_factory=_minicpmv_field_config,
340
341
            )

342
343
344
345
346
347
348
349
350
351
        return super()._parse_video_data(data)


class MiniCPMVProcessingInfo(BaseProcessingInfo):
    image_pattern = "(<image>./</image>)"
    video_pattern = "(<video>./</video>)"

    def get_hf_config(self):
        return self.ctx.get_hf_config()

352
353
    def get_hf_processor(self, **kwargs: object):
        hf_processor = self.ctx.get_hf_processor(**kwargs)
354
355
356
357
358
359
360
361
362

        # NumPy arrays are considered as Iterable but not Sequence in
        # https://github.com/huggingface/transformers/blob/main/src/transformers/image_transforms.py#L428
        image_processor = hf_processor.image_processor  # type: ignore
        for attr in ("mean", "std"):
            val = getattr(image_processor, attr)
            if isinstance(val, np.ndarray):
                setattr(image_processor, attr, val.tolist())

363
364
        return hf_processor

365
366
    def get_image_processor(self, **kwargs: object):
        return self.get_hf_processor(**kwargs).image_processor
367
368
369
370
371

    def get_model_version(self):
        return get_version_by_config(self.get_hf_config())

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
372
        mm_limits = {"image": None}
tc-mb's avatar
tc-mb committed
373
374
375
        if self.get_model_version() == (2,
                                        6) or self.get_model_version() == (4,
                                                                           0):
376
377
378
            mm_limits["video"] = None

        return mm_limits
379

380
381
382
383
384
385
386
387
388
389
    def get_slice_image_placeholder(
        self,
        image_size: ImageSize,
        # For MiniCPM V/O 2.6
        image_idx: int = 0,
        max_slice_nums: Optional[int] = None,
        use_image_id: bool = True,
    ) -> str:
        image_processor = self.get_image_processor()
        version = self.get_model_version()
390

391
392
        if version == (2, 0) or version == (2, 5):
            return image_processor.get_slice_image_placeholder(image_size)
393

394
395
396
397
398
399
        return image_processor.get_slice_image_placeholder(
            image_size,
            image_idx=image_idx,
            max_slice_nums=max_slice_nums,
            use_image_id=use_image_id,
        )
400

401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
    def get_sliced_grid(
        self,
        image_size: ImageSize,
        # For MiniCPM V/O 2.6
        max_slice_nums: Optional[int] = None,
    ) -> Optional[tuple[int, int]]:
        image_processor = self.get_image_processor()
        version = self.get_model_version()

        if version == (2, 0) or version == (2, 5):
            return image_processor.get_sliced_grid(image_size)

        if max_slice_nums is None:
            max_slice_nums = image_processor.max_slice_nums

        return image_processor.get_sliced_grid(
            image_size,
            max_slice_nums=max_slice_nums,
        )

421
422
423
424
425
    def get_num_image_tokens(
        self,
        image_size: ImageSize,
        max_slice_nums: Optional[int] = None,
    ) -> int:
426
427
428
        image_processor = self.get_image_processor()

        grid = self.get_sliced_grid(
429
430
431
            image_size,
            max_slice_nums=max_slice_nums,
        )
432
433
434
435
        if grid is None:
            ncols = nrows = 0
        else:
            ncols, nrows = grid
436

437
        return (ncols * nrows + 1) * image_processor.image_feature_size
438
439
440

    def get_max_image_tokens(self) -> int:
        image_size = self.get_image_size_with_most_features()
441
442
443
444
        return self.get_num_image_tokens(image_size)

    def get_image_max_slice_num(self) -> int:
        return getattr(self.get_hf_config(), "max_slice_num", 9)
445
446

    def get_image_size_with_most_features(self) -> ImageSize:
447
448
449
450
451
452
453
454
455
456
457
458
        image_size = getattr(self.get_hf_config(), "image_size", 448)
        max_slice_num = self.get_image_max_slice_num()
        return ImageSize(width=image_size, height=image_size * max_slice_num)

    def get_max_video_frame_tokens(self) -> int:
        frame_size = self.get_video_frame_size_with_most_features()

        return self.get_num_image_tokens(
            frame_size,
            max_slice_nums=self.get_video_max_slice_num(),
        )

459
460
461
462
463
464
465
466
    def get_max_video_tokens(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        num_frames = self.get_num_frames_with_most_features(seq_len, mm_counts)
        num_video_tokens_total = self.get_max_video_frame_tokens() * num_frames
        return num_video_tokens_total
467
468
469

    def get_video_max_slice_num(self) -> int:
        return 1
470

471
    def get_video_frame_size_with_most_features(self) -> ImageSize:
472
473
474
        image_size = getattr(self.get_hf_config(), "image_size", 448)
        max_slice_num = self.get_video_max_slice_num()
        return ImageSize(width=image_size, height=image_size * max_slice_num)
475

476
477
478
479
    def get_max_video_frames(self, max_tokens: int) -> int:
        num_frame_tokens = self.get_max_video_frame_tokens()
        num_frames = max_tokens // num_frame_tokens
        return num_frames
480

481
482
483
484
485
486
487
    def get_num_frames_with_most_features(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        max_images = mm_counts.get("image", 0)
        max_videos = mm_counts.get("video", 0)
488

489
        max_image_tokens = self.get_max_image_tokens() * max_images
490
491
        max_total_frames = self.get_max_video_frames(seq_len -
                                                     max_image_tokens)
492
493
        max_frames_per_video = min(max_total_frames // max(max_videos, 1),
                                   _MAX_FRAMES_PER_VIDEO)
494

495
        return max(max_frames_per_video, 1)
496
497


498
499
500
501
502
503
_I = TypeVar("_I",
             bound=MiniCPMVProcessingInfo,
             default=MiniCPMVProcessingInfo)


class MiniCPMVDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
504

505
506
507
508
509
510
511
512
513
514
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

        image_prompt_texts = self.info.image_pattern * num_images
        video_prompt_texts = self.info.video_pattern * num_videos

        return image_prompt_texts + video_prompt_texts

    def get_dummy_mm_data(
515
516
517
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
518
    ) -> MultiModalDataDict:
519
520
521
522
523
524
525
526
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

        image_width, image_height = \
            self.info.get_image_size_with_most_features()
        video_width, video_height = \
            self.info.get_video_frame_size_with_most_features()
        num_video_frames = \
527
            self.info.get_num_frames_with_most_features(seq_len, mm_counts)
528

529
        return {
530
531
532
533
534
535
536
537
538
539
540
541
            "image":
            self._get_dummy_images(width=image_width,
                                   height=image_height,
                                   num_images=num_images),
            "video": [
                self._get_dummy_images(width=video_width,
                                       height=video_height,
                                       num_images=num_video_frames)
            ] * num_videos,
        }


542
class MiniCPMVMultiModalProcessor(BaseMultiModalProcessor[_I]):
543
544
545
546
547
548
549

    def _get_data_parser(self) -> MultiModalDataParser:
        return MiniCPMVMultiModalDataParser()

    def get_image_prompt_texts(self,
                               image_size: ImageSize,
                               image_idx: int = 0) -> str:
550
551
552
553
        return self.info.get_slice_image_placeholder(
            image_size,
            image_idx=image_idx,
        )
554
555
556

    def get_video_prompt_texts(self, image_size: ImageSize,
                               num_frames: int) -> str:
557
        return self.info.get_slice_image_placeholder(
558
559
560
561
562
            image_size=image_size,
            image_idx=0,
            max_slice_nums=self.info.get_video_max_slice_num(),
            use_image_id=False,
        ) * num_frames
563

564
565
566
567
    def process_images(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
568
        tok_kwargs: Mapping[str, object],
569
    ) -> Mapping[str, NestedTensors]:
570
571
572
573
574
        if (images := mm_data.get("images")) is None:
            return {}

        parsed_images = (self._get_data_parser().parse_mm_data({
            "image": images
575
576
        }).get_items("image",
                     (MiniCPMVImageEmbeddingItems, ImageProcessorItems)))
577

578
579
580
581
582
583
584
        if isinstance(parsed_images, MiniCPMVImageEmbeddingItems):
            image_inputs = {}
        else:
            image_inputs = self._base_call_hf_processor(
                prompts=[self.info.image_pattern] * len(parsed_images),
                mm_data={"images": [[image] for image in parsed_images]},
                mm_kwargs=mm_kwargs,
585
                tok_kwargs=tok_kwargs,
586
587
588
589
590
591
592
593
                out_keys={"pixel_values", "image_sizes", "tgt_sizes"},
            )

        tokenizer = self.info.get_tokenizer()
        unk_token_id = tokenizer.get_vocab()["<unk>"]
        image_inputs["image_token_id"] = torch.tensor(unk_token_id)

        return image_inputs
594

595
596
597
598
    def process_videos(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
599
        tok_kwargs: Mapping[str, object],
600
    ) -> Mapping[str, NestedTensors]:
601
602
603
604
605
        if (videos := mm_data.get("videos")) is None:
            return {}

        parsed_videos = (self._get_data_parser().parse_mm_data({
            "video": videos
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
        }).get_items("video",
                     (MiniCPMVVideoEmbeddingItems, VideoProcessorItems)))

        if isinstance(parsed_videos, MiniCPMVVideoEmbeddingItems):
            video_inputs = {}
        else:
            video_inputs = self._base_call_hf_processor(
                prompts=[
                    self.info.image_pattern * len(video)
                    for video in parsed_videos
                ],
                mm_data={"images": list(parsed_videos)},
                mm_kwargs={
                    **mm_kwargs,
                    "max_slice_nums":
                    self.info.get_video_max_slice_num(),
                },
623
                tok_kwargs=tok_kwargs,
624
625
626
                out_keys={"pixel_values", "image_sizes", "tgt_sizes"},
            )

627
628
        video_inputs = {f"video_{k}": v for k, v in video_inputs.items()}

629
        tokenizer = self.info.get_tokenizer()
630
631
        unk_token_id = tokenizer.get_vocab()["<unk>"]
        video_inputs["video_token_id"] = torch.tensor(unk_token_id)
632

633
        return video_inputs
634

635
636
637
638
    def process_mm_inputs(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
639
        tok_kwargs: Mapping[str, object],
640
    ) -> Mapping[str, NestedTensors]:
641
        return {
642
643
            **self.process_images(mm_data, mm_kwargs, tok_kwargs),
            **self.process_videos(mm_data, mm_kwargs, tok_kwargs),
644
        }
645

646
    def _base_call_hf_processor(
647
        self,
648
649
        prompts: list[str],
        mm_data: Mapping[str, Sequence[object]],
650
        mm_kwargs: Mapping[str, object],
651
        tok_kwargs: Mapping[str, object],
652
653
        *,
        out_keys: set[str],
654
    ) -> dict[str, NestedTensors]:
655
        # This processor supports zipping prompt and mm_data together
tc-mb's avatar
tc-mb committed
656
657
        if self.info.get_model_version() == (
                2, 6) or self.info.get_model_version() == (4, 0):
658
659
660
661
            inputs = super()._call_hf_processor(
                prompt=prompts,  # type: ignore
                mm_data=mm_data,
                mm_kwargs=mm_kwargs,
662
                tok_kwargs=tok_kwargs,
663
664
665
666
667
668
669
670
671
672
673
674
            )
        else:
            inputs = defaultdict[str, list[torch.Tensor]](list)

            for i, prompt in enumerate(prompts):
                inputs_one = super()._call_hf_processor(
                    prompt=prompt,
                    mm_data={
                        k: v[i]
                        for k, v in mm_data.items()
                    },
                    mm_kwargs=mm_kwargs,
675
                    tok_kwargs=tok_kwargs,
676
677
678
679
680
681
682
                )

                for k, v in inputs_one.items():
                    assert len(v) == 1, (k, len(v))
                    inputs[k].append(v[0])

        return {k: inputs[k] for k in out_keys}
683
684
685
686
687
688

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
689
        tok_kwargs: Mapping[str, object],
690
691
    ) -> BatchFeature:
        tokenizer = self.info.get_tokenizer()
692

693
694
        input_ids = torch.tensor([tokenizer.encode(prompt, **tok_kwargs)])
        mm_inputs = self.process_mm_inputs(mm_data, mm_kwargs, tok_kwargs)
695
696

        return BatchFeature({
697
            "input_ids": input_ids,
698
            **mm_inputs,
699
        })
700

701
    def _hf_processor_applies_updates(
702
703
704
705
        self,
        prompt_text: str,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
706
        tokenization_kwargs: Mapping[str, object],
707
708
709
    ) -> bool:
        return False

710
    def _get_prompt_updates(
711
712
713
714
715
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> Sequence[PromptUpdate]:
tc-mb's avatar
tc-mb committed
716
717
718
719
720
721
722
723
724
725
726
727
        placeholders = [("image", self.info.image_pattern),
                        ("video", self.info.video_pattern)]

        # hard code for inconsistency of encode-decode image_pattern
        additional_placeholders = []
        tokenizer = self.info.get_tokenizer()
        for modality, pattern in placeholders:
            sub_pattern = tokenizer.decode(
                tokenizer.encode(pattern, add_special_tokens=False))
            if sub_pattern != pattern:
                additional_placeholders.append((modality, sub_pattern))
        placeholders += additional_placeholders
728

729
730
731
732
733
734
        def get_image_replacement(item_idx: int):
            images = mm_items.get_items(
                "image", (MiniCPMVImageEmbeddingItems, ImageProcessorItems))

            image_size = images.get_image_size(item_idx)

735
736
737
738
            return PromptUpdateDetails.select_text(
                self.get_image_prompt_texts(image_size, item_idx),
                "<unk>",
            )
739
740
741
742
743
744
745
746

        def get_video_replacement(item_idx: int):
            videos = mm_items.get_items(
                "video", (MiniCPMVVideoEmbeddingItems, VideoProcessorItems))

            frame_size = videos.get_frame_size(item_idx)
            num_frames = videos.get_num_frames(item_idx)

747
748
749
750
            return PromptUpdateDetails.select_text(
                self.get_video_prompt_texts(frame_size, num_frames),
                "<unk>",
            )
751
752
753
754
755

        get_replacement = {
            "image": get_image_replacement,
            "video": get_video_replacement,
        }
756
757
758

        return [
            PromptReplacement(modality=modality,
tc-mb's avatar
tc-mb committed
759
                              target=pattern,
760
                              replacement=get_replacement[modality])
tc-mb's avatar
tc-mb committed
761
            for modality, pattern in placeholders
762
        ]
763

764
765
    def _get_mm_fields_config(
        self,
766
        hf_inputs: BatchFeature,
767
768
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
769
        return _minicpmv_field_config(hf_inputs)
770

771
772

class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
Jee Jee Li's avatar
Jee Jee Li committed
773
774
775
776
    """
    The abstract class of MiniCPMV can only be inherited, but cannot be
    instantiated.
    """
777

778
779
780
781
782
783
784
785
786
    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "(<image>./</image>)"
        if modality.startswith("video"):
            return "(<video>./</video>)"

        raise ValueError("Only image or video modality is supported")

787
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
788
789
790
        config = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config
        quant_config = vllm_config.quant_config
791
        super().__init__()
792
793
794
795
        # All MiniCPM-V models disable `tie_word_embeddings` but
        # `PretrainedConfig.tie_word_embeddings` defaults to True; we cannot
        # check `tie_word_embeddings` until vLLM integrate MiniCPM-V model
        # and config class
796
797
798
        self.config = config
        self.multimodal_config = multimodal_config

799
        self.version = get_version_by_config(self.config)
800
801
802
803
804
        self.llm = self.init_llm(vllm_config=vllm_config,
                                 prefix=maybe_prefix(prefix, "llm"))
        self.vpm = self.init_vision_module(config,
                                           quant_config,
                                           prefix=maybe_prefix(prefix, "vpm"))
Jee Jee Li's avatar
Jee Jee Li committed
805
806
        self.vision_dim = (self.vpm.embed_dim if self.version == (2, 0) else
                           self.vpm.embeddings.embed_dim)
Alphi's avatar
Alphi committed
807
        self.embed_dim = self.config.hidden_size
808

809
810
811
        self.resampler = self.init_resampler(self.embed_dim,
                                             self.vision_dim,
                                             quant_config=quant_config,
812
813
                                             prefix=maybe_prefix(
                                                 prefix, "resampler"))
814

815
        self.mm_token_ids = set[int]()
816
817
818
        self.make_empty_intermediate_tensors = (
            self.llm.make_empty_intermediate_tensors)

819
    def _parse_and_validate_vision_input(
Jee Jee Li's avatar
Jee Jee Li committed
820
        self,
821
        modality: str,
Jee Jee Li's avatar
Jee Jee Li committed
822
        **kwargs: object,
823
    ) -> Optional[MiniCPMVImageInputs]:
824
825
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)
826

827
        if pixel_values is None and image_embeds is None:
828
829
            return None

830
831
832
833
834
835
836
837
838
839
840
841
        image_token_id = kwargs.pop("image_token_id")
        if image_token_id is not None:
            assert isinstance(image_token_id, torch.Tensor)
            self.mm_token_ids.add(image_token_id.flatten().unique().item())

        if image_embeds is not None:
            if not isinstance(image_embeds, (torch.Tensor, list)):
                raise ValueError(
                    f"Incorrect type of image_embeds for {modality=}. "
                    f"Got type: {type(image_embeds)}")

            image_embeds_flat = flatten_bn(image_embeds)
842

843
            return MiniCPMVImageEmbeddingInputs(
844
                type="image_embeds",
845
                image_embeds=image_embeds_flat,
846
            )
847

848
849
850
851
        if not isinstance(pixel_values, (torch.Tensor, list)):
            raise ValueError(
                f"Incorrect type of pixel_values for {modality=}. "
                f"Got type: {type(pixel_values)}")
852

853
854
855
856
857
858
859
860
861
862
        tgt_sizes = kwargs.pop("tgt_sizes")
        if not isinstance(tgt_sizes, (torch.Tensor, list)):
            raise ValueError(f"Incorrect type of tgt_sizes for {modality=}. "
                             f"Got type: {type(tgt_sizes)}")

        num_slices = [[len(p) for p in ps] for ps in pixel_values]
        num_slices_flat = flatten_bn(torch.tensor(num_slices))

        pixel_values_flat = flatten_bn(flatten_2d_lists(pixel_values))
        tgt_sizes_flat = flatten_bn(flatten_2d_lists(tgt_sizes), concat=True)
863

864
        return MiniCPMVImagePixelInputs(
865
866
            type="pixel_values",
            pixel_values=pixel_values_flat,
867
868
            tgt_sizes=tgt_sizes_flat,
            num_slices=num_slices_flat,
Jee Jee Li's avatar
Jee Jee Li committed
869
        )
870

871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
        modalities = {}

        # Preserve the order of modalities if there are multiple of them
        # from the order of kwargs.
        for input_key in kwargs:
            if input_key in ("pixel_values",
                             "image_embeds") and "images" not in modalities:
                modalities["images"] = self._parse_and_validate_vision_input(
                    "images", **kwargs)
            if input_key in ("video_pixel_values",
                             "video_embeds") and "videos" not in modalities:

                def _image_key(video_key: str):
                    if video_key == "video_token_id":
                        return "image_token_id"

                    return video_key.removeprefix("video_")

                modalities["videos"] = self._parse_and_validate_vision_input(
                    "videos", **{
                        _image_key(k): v
                        for k, v in kwargs.items()
                    })

        return modalities

    def _process_vision_input(
        self,
        image_input: MiniCPMVImageInputs,
    ) -> Union[torch.Tensor, list[torch.Tensor], tuple[torch.Tensor, ...]]:
        if image_input["type"] == "image_embeds":
            return image_input["image_embeds"]

        image_features_flat = self.get_vision_hidden_states(image_input)

907
908
909
910
911
        num_slices = image_input["num_slices"]
        return [
            e.flatten(0, 1)
            for e in image_features_flat.split(num_slices.tolist())
        ]
912
913
914
915
916
917
918
919
920
921
922
923

    def _process_multimodal_inputs(self, modalities: dict):
        # The result multimodal_embeddings is tuple of tensors, with each
        # tensor correspoending to a multimodal data item (image or video).
        multimodal_embeddings: tuple[torch.Tensor, ...] = ()

        # NOTE: It is important to iterate over the keys in this dictionary
        # to preserve the order of the modalities.
        for modality in modalities:
            if modality == "images":
                image_input = modalities["images"]
                image_features = self._process_vision_input(image_input)
924
                multimodal_embeddings += tuple(image_features)
925
926
927
            if modality == "videos":
                video_input = modalities["videos"]
                video_features = self._process_vision_input(video_input)
928
                multimodal_embeddings += tuple(video_features)
929
930
931

        return multimodal_embeddings

932
933
934
    def get_language_model(self) -> torch.nn.Module:
        return self.llm

935
936
    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
937
938
        modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
        if not modalities:
939
            return []
940
941
942
943
944
945
946
947
948

        return self._process_multimodal_inputs(modalities)

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
    ) -> torch.Tensor:
        inputs_embeds = self.llm.get_input_embeddings(input_ids)
949
950
        if multimodal_embeddings is not None \
            and len(multimodal_embeddings) != 0:
951
952
953
954
            assert len(self.mm_token_ids) > 0
            inputs_embeds = merge_multimodal_embeddings(
                input_ids,
                inputs_embeds,
955
                multimodal_embeddings,
956
957
958
                list(self.mm_token_ids),
            )
        return inputs_embeds
959

960
961
962
963
964
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
965
        inputs_embeds: Optional[torch.Tensor] = None,
Jee Jee Li's avatar
Jee Jee Li committed
966
967
        **kwargs: Any,
    ) -> torch.Tensor:
968
        if intermediate_tensors is not None:
969
970
971
972
973
974
975
            inputs_embeds = None

        # NOTE: In v1, inputs_embeds is always generated at model runner from
        # `get_multimodal_embeddings` and `get_input_embeddings`, this
        # condition is only for v0 compatibility.
        elif inputs_embeds is None:
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
Jee Jee Li's avatar
Jee Jee Li committed
976

977
978
979
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None
980

981
        hidden_states = self.llm.model(
982
            input_ids=input_ids,
Jee Jee Li's avatar
Jee Jee Li committed
983
984
            positions=positions,
            intermediate_tensors=intermediate_tensors,
985
            inputs_embeds=inputs_embeds,
Jee Jee Li's avatar
Jee Jee Li committed
986
        )
987
        return hidden_states
988

989
990
991
992
993
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
994
        return self.llm.compute_logits(hidden_states, sampling_metadata)
995

996
997
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
998
999
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)
Jee Jee Li's avatar
Jee Jee Li committed
1000

1001
1002
1003
1004
1005
1006
1007
1008
    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(language_model="llm",
                                                connector="resampler",
                                                tower_model="vpm")

Jee Jee Li's avatar
Jee Jee Li committed
1009
1010
    def init_llm(
        self,
1011
        vllm_config: VllmConfig,
1012
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1013
1014
1015
    ) -> nn.Module:
        raise NotImplementedError

1016
1017
1018
1019
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1020
        prefix: str = "",
1021
    ) -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1022
1023
        raise NotImplementedError

1024
1025
1026
1027
1028
    def init_resampler(self,
                       embed_dim: int,
                       vision_dim: int,
                       quant_config: Optional[QuantizationConfig] = None,
                       prefix: str = "") -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1029
1030
        raise NotImplementedError

1031
1032
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
Jee Jee Li's avatar
Jee Jee Li committed
1033
1034
1035
        raise NotImplementedError


1036
class MiniCPMV2_0(MiniCPMVBaseModel):
Jee Jee Li's avatar
Jee Jee Li committed
1037

1038
1039
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1040
1041
1042
1043
        assert self.version == (2, 0)

    def init_llm(
        self,
1044
        vllm_config: VllmConfig,
1045
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1046
    ) -> nn.Module:
1047
        return MiniCPMForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1048

1049
1050
1051
1052
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1053
        prefix: str = "",
1054
    ) -> nn.Module:
1055
        # TODO: refactor vision model through timm wrapper from transformers
Jee Jee Li's avatar
Jee Jee Li committed
1056
1057
1058
1059
        try:
            import timm
        except ImportError:
            raise ImportError("Please install timm==0.9.10") from ImportError
1060

Jee Jee Li's avatar
Jee Jee Li committed
1061
1062
1063
1064
1065
1066
1067
1068
1069
        with set_default_torch_dtype(torch.float16):
            model = timm.create_model(
                "vit_so400m_patch14_siglip_384.webli",
                pretrained=False,
                num_classes=0,
                dynamic_img_size=True,
                dynamic_img_pad=True,
            )

1070
1071
        model = model.to(dtype=torch.get_default_dtype())

Jee Jee Li's avatar
Jee Jee Li committed
1072
1073
1074
1075
1076
1077
1078
1079
1080
        if (isinstance(model, timm.models.VisionTransformer)
                and model.attn_pool is not None):
            model.attn_pool = torch.nn.Identity()

        if self.config.drop_vision_last_layer:
            model.blocks = model.blocks[:-1]

        return model

1081
1082
1083
1084
1085
    def init_resampler(self,
                       embed_dim: int,
                       vision_dim: int,
                       quant_config: Optional[QuantizationConfig] = None,
                       prefix: str = "") -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1086
        with set_default_torch_dtype(torch.float16):
1087
1088
1089
1090
1091
1092
1093
1094
1095
            resampler = Resampler2(embed_dim=embed_dim,
                                   num_heads=embed_dim // 128,
                                   grid_size=int(
                                       math.sqrt(self.config.query_num)),
                                   kv_dim=vision_dim,
                                   adaptive=False,
                                   do_post_projection=True,
                                   quant_config=quant_config,
                                   prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1096

1097
1098
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1099

1100
1101
1102
1103
1104
1105
1106
1107
1108
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]

        P_h, P_w = self.vpm.patch_embed.patch_size
        dtype: torch.dtype = self.vpm.pos_embed.data.dtype
        num_prefix_tokens = getattr(self.vpm, "num_prefix_tokens", 0)

        res = list[torch.Tensor]()
Jee Jee Li's avatar
Jee Jee Li committed
1109
1110
        for pixel_value in pixel_values:
            H, W = pixel_value[0].shape[-2:]
1111
            tgt_size = (math.ceil(H / P_h), math.ceil(W / P_w))
Jee Jee Li's avatar
Jee Jee Li committed
1112
1113
1114
            vision_embedding = self.vpm.forward_features(
                pixel_value.unsqueeze(0).type(dtype))

1115
1116
1117
            if num_prefix_tokens > 0:
                vision_embedding = vision_embedding[:, num_prefix_tokens:]
            res.append(self.resampler(vision_embedding, tgt_size))
Jee Jee Li's avatar
Jee Jee Li committed
1118

1119
        return torch.vstack(res)
Jee Jee Li's avatar
Jee Jee Li committed
1120
1121


1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
Jee Jee Li's avatar
Jee Jee Li committed
1134

1135
1136
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1137
1138
1139
1140
        assert self.version == (2, 5)

    def init_llm(
        self,
1141
        vllm_config: VllmConfig,
1142
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1143
    ) -> nn.Module:
1144
        return LlamaForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1145

1146
1147
1148
1149
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1150
        prefix: str = "",
1151
1152
    ) -> nn.Module:
        model = Idefics2VisionTransformer(config.vision_config,
1153
1154
                                          quant_config=quant_config,
                                          prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1155
1156
1157
1158
        if self.config.drop_vision_last_layer:
            model.encoder.layers = model.encoder.layers[:-1]
        return model

1159
1160
1161
1162
1163
    def init_resampler(self,
                       embed_dim: int,
                       vision_dim: int,
                       quant_config: Optional[QuantizationConfig] = None,
                       prefix: str = "") -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1164
        with set_default_torch_dtype(torch.float16):
1165
1166
1167
1168
1169
1170
            resampler = Resampler2_5(num_queries=self.config.query_num,
                                     embed_dim=embed_dim,
                                     num_heads=embed_dim // 128,
                                     kv_dim=vision_dim,
                                     quant_config=quant_config,
                                     prefix=prefix)
1171

1172
1173
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1174

1175
1176
1177
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]
Jee Jee Li's avatar
Jee Jee Li committed
1178
1179
        tgt_sizes = data["tgt_sizes"]

1180
1181
1182
1183
1184
        B = len(pixel_values)
        P = pixel_values[0].shape[-2]
        L = max(item.shape[-1] for item in pixel_values)
        device = pixel_values[0].device
        dtype = pixel_values[0].dtype
Jee Jee Li's avatar
Jee Jee Li committed
1185

1186
1187
1188
1189
1190
1191
        all_pixel_values = torch.zeros((B, 3, P, L),
                                       dtype=dtype,
                                       device=device)
        for i, pixel_values_item in enumerate(pixel_values):
            L_item = pixel_values_item.shape[-1]
            all_pixel_values[i, ..., :L_item] = pixel_values_item
Jee Jee Li's avatar
Jee Jee Li committed
1192

1193
1194
1195
        num_patches = tgt_sizes.prod(-1)
        max_patches = num_patches.max().item()
        assert isinstance(max_patches, int)
Jee Jee Li's avatar
Jee Jee Li committed
1196

1197
        patch_attn_mask = torch.zeros((B, max_patches),
Jee Jee Li's avatar
Jee Jee Li committed
1198
1199
                                      dtype=torch.bool,
                                      device=device)
1200
1201
        for i, num_patches_item in enumerate(num_patches):
            patch_attn_mask[i, :num_patches_item] = True
Jee Jee Li's avatar
Jee Jee Li committed
1202

1203
1204
1205
1206
1207
1208
1209
        vision_embedding = self.vpm(
            all_pixel_values,
            patch_attention_mask=patch_attn_mask.unsqueeze(1),
            tgt_sizes=None,
        )

        return self.resampler(vision_embedding, tgt_sizes)
Jee Jee Li's avatar
Jee Jee Li committed
1210
1211


1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
Jee Jee Li's avatar
Jee Jee Li committed
1224

1225
1226
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
1227
        assert self.version == (2, 6)
Jee Jee Li's avatar
Jee Jee Li committed
1228
1229
1230

    def init_llm(
        self,
1231
        vllm_config: VllmConfig,
1232
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1233
    ) -> nn.Module:
1234
        return Qwen2ForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1235

1236
1237
1238
    def init_vision_module(
        self,
        config: PretrainedConfig,
1239
        quant_config: Optional[QuantizationConfig] = None,
1240
        prefix: str = "",
1241
1242
    ) -> nn.Module:
        model = Idefics2VisionTransformer(config.vision_config,
1243
1244
                                          quant_config=quant_config,
                                          prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1245
1246
1247
1248
        if self.config.drop_vision_last_layer:
            model.encoder.layers = model.encoder.layers[:-1]
        return model

1249
1250
1251
1252
1253
    def init_resampler(self,
                       embed_dim: int,
                       vision_dim: int,
                       quant_config: Optional[QuantizationConfig] = None,
                       prefix: str = "") -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1254
        with set_default_torch_dtype(torch.float16):
1255
            # The resampler in 2.6 remains consistent with the one in 2.5.
1256
1257
1258
1259
1260
1261
            resampler = Resampler2_5(num_queries=self.config.query_num,
                                     embed_dim=embed_dim,
                                     num_heads=embed_dim // 128,
                                     kv_dim=vision_dim,
                                     quant_config=quant_config,
                                     prefix=prefix)
1262

1263
1264
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1265

1266
1267
1268
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]
Jee Jee Li's avatar
Jee Jee Li committed
1269
1270
        tgt_sizes = data["tgt_sizes"]

1271
1272
1273
1274
1275
        B = len(pixel_values)
        P = pixel_values[0].shape[-2]
        L = max(item.shape[-1] for item in pixel_values)
        device = pixel_values[0].device
        dtype = pixel_values[0].dtype
Jee Jee Li's avatar
Jee Jee Li committed
1276

1277
1278
1279
1280
1281
1282
        all_pixel_values = torch.zeros((B, 3, P, L),
                                       dtype=dtype,
                                       device=device)
        for i, pixel_values_item in enumerate(pixel_values):
            L_item = pixel_values_item.shape[-1]
            all_pixel_values[i, ..., :L_item] = pixel_values_item
Jee Jee Li's avatar
Jee Jee Li committed
1283

1284
1285
1286
        num_patches = tgt_sizes.prod(-1)
        max_patches = num_patches.max().item()
        assert isinstance(max_patches, int)
Jee Jee Li's avatar
Jee Jee Li committed
1287

1288
        patch_attn_mask = torch.zeros((B, max_patches),
Jee Jee Li's avatar
Jee Jee Li committed
1289
1290
                                      dtype=torch.bool,
                                      device=device)
1291
1292
1293
        for i, num_patches_item in enumerate(num_patches):
            patch_attn_mask[i, :num_patches_item] = True

Jee Jee Li's avatar
Jee Jee Li committed
1294
        vision_embedding = self.vpm(
1295
1296
            all_pixel_values,
            patch_attention_mask=patch_attn_mask.unsqueeze(1),
Jee Jee Li's avatar
Jee Jee Li committed
1297
            tgt_sizes=tgt_sizes,
1298
        )
Jee Jee Li's avatar
Jee Jee Li committed
1299
1300
1301

        return self.resampler(vision_embedding, tgt_sizes)

tc-mb's avatar
tc-mb committed
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self,
                                   skip_prefixes=["apm.", "audio", "tts"])
        return loader.load_weights(weights)


class MiniCPMV4_0(MiniCPMVBaseModel, SupportsLoRA):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
        assert self.version == (4, 0)

    def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
        if isinstance(quant_config, (AWQConfig, AWQMarlinConfig)):
            return None
        return quant_config

    def init_llm(
        self,
        vllm_config: VllmConfig,
        prefix: str = "",
    ) -> nn.Module:
        return LlamaForCausalLM(vllm_config=vllm_config, prefix=prefix)

    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> nn.Module:
        quant_config = self._maybe_ignore_quant_config(quant_config)
        model = Idefics2VisionTransformer(config.vision_config,
                                          quant_config=quant_config,
                                          prefix=prefix)
        if self.config.drop_vision_last_layer:
            model.encoder.layers = model.encoder.layers[:-1]
        return model

    def init_resampler(
        self,
        embed_dim: int,
        vision_dim: int,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> nn.Module:
        quant_config = self._maybe_ignore_quant_config(quant_config)
        with set_default_torch_dtype(torch.float16):
            # The resampler in 4.0 remains consistent with the one in 2.5/2.6.
            resampler = Resampler2_5(num_queries=self.config.query_num,
                                     embed_dim=embed_dim,
                                     num_heads=embed_dim // 128,
                                     kv_dim=vision_dim,
                                     quant_config=quant_config,
                                     prefix=prefix)

        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())

    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]
        tgt_sizes = data["tgt_sizes"]

        B = len(pixel_values)
        P = pixel_values[0].shape[-2]
        L = max(item.shape[-1] for item in pixel_values)
        device = pixel_values[0].device
        dtype = pixel_values[0].dtype

        all_pixel_values = torch.zeros((B, 3, P, L),
                                       dtype=dtype,
                                       device=device)
        for i, pixel_values_item in enumerate(pixel_values):
            L_item = pixel_values_item.shape[-1]
            all_pixel_values[i, ..., :L_item] = pixel_values_item

        num_patches = tgt_sizes.prod(-1)
        max_patches = num_patches.max().item()
        assert isinstance(max_patches, int)

        patch_attn_mask = torch.zeros((B, max_patches),
                                      dtype=torch.bool,
                                      device=device)
        for i, num_patches_item in enumerate(num_patches):
            patch_attn_mask[i, :num_patches_item] = True

        vision_embedding = self.vpm(
            all_pixel_values,
            patch_attention_mask=patch_attn_mask.unsqueeze(1),
            tgt_sizes=tgt_sizes,
        )

        return self.resampler(vision_embedding, tgt_sizes)

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self,
                                   skip_prefixes=["apm.", "audio", "tts"])
        return loader.load_weights(weights)

Jee Jee Li's avatar
Jee Jee Li committed
1414

1415
1416
1417
_SUPPORT_VERSION = {
    (2, 0): MiniCPMV2_0,
    (2, 5): MiniCPMV2_5,
1418
    (2, 6): MiniCPMV2_6,
tc-mb's avatar
tc-mb committed
1419
    (4, 0): MiniCPMV4_0,
1420
1421
1422
}


1423
1424
1425
1426
1427
@MULTIMODAL_REGISTRY.register_processor(
    MiniCPMVMultiModalProcessor,
    info=MiniCPMVProcessingInfo,
    dummy_inputs=MiniCPMVDummyInputsBuilder)
class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
Jee Jee Li's avatar
Jee Jee Li committed
1428
1429
1430
1431
1432
    """
    Different versions of MiniCPMV use different visual encoders and LLMs,
    which is not conducive to the current integration logic of LoRA and
    bitsandbytes in vLLM. Therefore, it is necessary to separate them.
    """
1433

1434
    def __new__(cls, *, vllm_config: VllmConfig, prefix: str = ""):
1435
        config = vllm_config.model_config.hf_config
Jee Jee Li's avatar
Jee Jee Li committed
1436
1437
1438
1439
1440
1441
1442
1443
1444
        if not hasattr(config, "version"):
            if config.hidden_size == 2304 and config.query_num == 64:
                version = (2, 0)
            else:
                version = (2, 5)
        else:
            version = str(config.version).split(".")
            version = tuple([int(x) for x in version])
        # Dispatch class based on version
1445
1446
        instance_cls = _SUPPORT_VERSION.get(version)
        if instance_cls is None:
tc-mb's avatar
tc-mb committed
1447
1448
1449
1450
            supported_versions = ", ".join(
                [f"{v[0]}.{v[1]}" for v in sorted(_SUPPORT_VERSION.keys())])
            raise ValueError(f"Currently, MiniCPMV only supports versions "
                             f"{supported_versions}. Got version: {version}")
1451
1452
1453
1454
1455
1456
1457

        # quant_config references base class members,
        # so update values before init is called
        cls.packed_modules_mapping.update(instance_cls.packed_modules_mapping)
        cls.embedding_modules.update(instance_cls.embedding_modules)
        cls.embedding_padding_modules += instance_cls.embedding_padding_modules
        return instance_cls(vllm_config=vllm_config, prefix=prefix)