minicpmv.py 63.8 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
tc-mb's avatar
tc-mb committed
30
from itertools import chain
31
from typing import Annotated, Any, Callable, Literal, Optional, Union
32

33
import numpy as np
34
import torch
Alphi's avatar
Alphi committed
35
import torch.types
36
from torch import nn
tc-mb's avatar
tc-mb committed
37
from torch.nn.init import trunc_normal_
38
from transformers import BatchFeature, PretrainedConfig
39
from typing_extensions import TypeVar
40

41
from vllm.config import VllmConfig
42
from vllm.config.multimodal import BaseDummyOptions
43
from vllm.model_executor.layers.quantization import QuantizationConfig
tc-mb's avatar
tc-mb committed
44
45
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
46
from vllm.model_executor.layers.resampler import (BaseResampler, Resampler2,
47
                                                  get_2d_sincos_pos_embed)
Jee Jee Li's avatar
Jee Jee Li committed
48
from vllm.model_executor.model_loader.utils import set_default_torch_dtype
49
50
from vllm.model_executor.models.llama import LlamaForCausalLM
from vllm.model_executor.models.minicpm import MiniCPMForCausalLM
51
from vllm.model_executor.models.module_mapping import MultiModelKeys
52
from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM
tc-mb's avatar
tc-mb committed
53
54
from vllm.model_executor.models.qwen3 import Qwen3ForCausalLM
from vllm.multimodal import MULTIMODAL_REGISTRY
55
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
tc-mb's avatar
tc-mb committed
56
                                    MultiModalKwargsItems, NestedTensors)
57
58
from vllm.multimodal.parse import (DictEmbeddingItems, ImageItem,
                                   ImageProcessorItems, ImageSize,
59
60
                                   ModalityData, ModalityDataItems,
                                   MultiModalDataItems, MultiModalDataParser,
61
                                   VideoItem, VideoProcessorItems)
62
from vllm.multimodal.processing import (BaseMultiModalProcessor,
63
                                        BaseProcessingInfo, PromptReplacement,
64
65
                                        PromptUpdate, PromptUpdateDetails,
                                        ResolvedPromptUpdate, _seq2text)
66
from vllm.multimodal.profiling import BaseDummyInputsBuilder
67
from vllm.platforms import current_platform
68
from vllm.sequence import IntermediateTensors
69
from vllm.utils import flatten_2d_lists
70
from vllm.utils.tensor_schema import TensorSchema, TensorShape
71

Jee Jee Li's avatar
Jee Jee Li committed
72
from .idefics2_vision_model import Idefics2VisionTransformer
73
74
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
                         SupportsMultiModal, SupportsPP)
75
from .utils import AutoWeightsLoader, flatten_bn, maybe_prefix
76

77
78
79
# For profile run
_MAX_FRAMES_PER_VIDEO = 16

80

81
class MiniCPMVImagePixelInputs(TensorSchema):
Jee Jee Li's avatar
Jee Jee Li committed
82
    """
83
84
85
86
87
88
    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
89
90
    """

91
92
93
94
95
96
    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],
97
        TensorShape("bns", "c", "h", "w", dynamic_dims={"h", "w"}),
98
99
100
101
102
103
104
105
106
107
108
109
    ]
    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
110
    """
111
112
113
114
    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
115
116
    """

117
    type: Literal["image_embeds"]
118
119
120
121
    image_embeds: Annotated[
        Union[torch.Tensor, list[torch.Tensor]],
        TensorShape("bn", "ns", "hs"),
    ]
122
123
124
125
126


MiniCPMVImageInputs = Union[MiniCPMVImagePixelInputs,
                            MiniCPMVImageEmbeddingInputs]

Jee Jee Li's avatar
Jee Jee Li committed
127
128
129
130
131
DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)


class Resampler2_5(BaseResampler):

132
133
134
135
136
137
    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,
138
                 max_size: tuple[int, int] = (70, 70),
139
140
141
142
143
144
145
146
147
                 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
148
149
150

        self.max_size = max_size
        self._set_2d_pos_cache(self.max_size)
151

Alphi's avatar
Alphi committed
152
    def _set_2d_pos_cache(self,
153
                          max_size: tuple[int, int],
Jee Jee Li's avatar
Jee Jee Li committed
154
155
156
157
158
                          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)
159
160
        self.register_buffer("pos_embed", pos_embed, persistent=False)

Alphi's avatar
Alphi committed
161
    def _adjust_pos_cache(self, tgt_sizes: torch.Tensor,
Jee Jee Li's avatar
Jee Jee Li committed
162
163
164
165
166
                          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)

167
        if max_h > self.max_size[0] or max_w > self.max_size[1]:
Jee Jee Li's avatar
Jee Jee Li committed
168
            self.max_size = (
169
                max(max_h, self.max_size[0]),
Jee Jee Li's avatar
Jee Jee Li committed
170
171
                max(max_w, self.max_size[1]),
            )
172
173
            self._set_2d_pos_cache(self.max_size, device)

Jee Jee Li's avatar
Jee Jee Li committed
174
175
    def forward(self, x: torch.Tensor,
                tgt_sizes: torch.Tensor) -> torch.Tensor:
176
177
178
179
180
181
182
183
184
185
        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
186
187
188
        max_patch_len = patch_len.max().item()
        assert isinstance(max_patch_len, int)

189
190
191
192
193
194
        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
195
            tgt_h, tgt_w = tgt_sizes[i].tolist()
196
197
198
199
200
201
202
203
            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
204
        x, _ = self.kv_proj(x)  # B * L * D
205
206
207
208
209
210
211
212
        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
213
214
            key_padding_mask=key_padding_mask,
        )[0]
215
216
217
218
219
220
221
222
        #  out: Q * B * D
        x = out.permute(1, 0, 2)  # B * Q * D

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


tc-mb's avatar
tc-mb committed
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
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
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
class Resampler4_5(Resampler2_5):

    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,
                 max_size: tuple[int, int] = (70, 70),
                 max_temporal_size: int = 36000,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = "") -> None:
        super().__init__(num_queries,
                         embed_dim,
                         num_heads,
                         kv_dim,
                         norm_layer,
                         max_size,
                         quant_config=quant_config,
                         prefix=prefix)

        trunc_normal_(self.query, std=.02)
        self.max_temporal_size = max_temporal_size
        self._set_temporal_pos_cache(self.max_temporal_size)
        self.apply(self._init_weights)

    def get_1d_sincos_pos_embed_from_temporal_size(self, embed_dim: int,
                                                   pos: np.ndarray):
        """
        embed_dim: output dimension for each position
        pos: a list of positions to be encoded: size (M,)
        out: (M, D)
        """
        assert embed_dim % 2 == 0
        omega = np.arange(embed_dim // 2, dtype=np.float32)
        omega /= embed_dim / 2.
        omega = 1. / 10000**omega  # (D/2,)

        pos = pos.reshape(-1)  # (M,)
        out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

        emb_sin = np.sin(out)  # (M, D/2)
        emb_cos = np.cos(out)  # (M, D/2)

        emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
        return emb

    def _set_temporal_pos_cache(self,
                                max_temporal_size: int,
                                device: torch.types.Device = "cpu") -> None:
        temporal_size = np.arange(max_temporal_size, dtype=np.float32)
        pos_embed = torch.from_numpy(
            self.get_1d_sincos_pos_embed_from_temporal_size(
                self.embed_dim, temporal_size)).float().to(device)
        self.register_buffer("temporal_pos_embed", pos_embed, persistent=False)

    def _adjust_temporal_pos_cache(self,
                                   max_temporal_size: int,
                                   device: torch.types.Device = "cpu"):
        if max_temporal_size > self.max_temporal_size:
            self.max_temporal_size = max_temporal_size
            self._set_temporal_pos_cache(self.max_temporal_size, device)

    def _init_weights(self, m: Union[nn.Linear, nn.LayerNorm]):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def forward(
        self,
        x: torch.Tensor,
        tgt_sizes: torch.Tensor,
        # temporal_ids for high refresh rate videos
        temporal_ids=None
    ) -> torch.Tensor:
        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)

        temporal_pos_emb = False
        temporal_ids_flatten = None
        if temporal_ids is not None:
            # example: [[-1], [-1], [2, 6, 9]]
            temporal_ids_flatten = list(chain.from_iterable(temporal_ids))
            max_temporal_size = max(temporal_ids_flatten, default=0)
            if max_temporal_size > -1:
                temporal_pos_emb = True
            if max_temporal_size > self.max_temporal_size:
                self._adjust_temporal_pos_cache(max_temporal_size, device)

        max_patch_len = patch_len.max().item()
        assert isinstance(max_patch_len, int)

        key_padding_mask = torch.zeros((bs, max_patch_len),
                                       dtype=torch.bool,
                                       device=device)

        x, _ = self.kv_proj(x)  # B * L * D
        x = self.ln_kv(x).permute(1, 0, 2)  # L * B * D
        q = self.ln_q(self.query)  # Q * D

        pos_embed_2d = []
        pos_embed_temporal = []
        for i in range(bs):
            tgt_h, tgt_w = tgt_sizes[i]
            if temporal_pos_emb:
                if temporal_ids_flatten[i] == -1:
                    pos_embed_temporal.append(
                        torch.zeros(self.embed_dim, dtype=dtype,
                                    device=device))
                else:
                    pos_embed_temporal.append(self.temporal_pos_embed[
                        temporal_ids_flatten[i]].to(dtype))  # D

            pos_embed_2d.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_2d = torch.nn.utils.rnn.pad_sequence(
            pos_embed_2d, batch_first=True,
            padding_value=0.0).permute(1, 0, 2)  # BLD => L * B * D

        k = x
        v = x + pos_embed_2d
        if pos_embed_temporal:
            k += torch.stack(pos_embed_temporal, dim=0)
            bs = len(temporal_ids)
            merge_k = []
            merge_v = []
            merge_key_padding_mask = []

            start = 0
            for tp in temporal_ids:
                end = start + len(tp)
                # L * (end-start) * D -> (end-start) * L * D
                # -> 1 * L*(end-start) * D
                merge_k.append(k[:, start:end, :].permute(1, 0, 2).reshape(
                    -1, self.embed_dim))
                merge_v.append(v[:, start:end, :].permute(1, 0, 2).reshape(
                    -1, self.embed_dim))
                merge_key_padding_mask.append(
                    key_padding_mask[start:end, :].reshape(-1, 1))

                start = end

            k = torch.nn.utils.rnn.pad_sequence(merge_k,
                                                batch_first=True,
                                                padding_value=0.0).permute(
                                                    1, 0, 2)  # L*(end-start)
            v = torch.nn.utils.rnn.pad_sequence(merge_v,
                                                batch_first=True,
                                                padding_value=0.0).permute(
                                                    1, 0, 2)  # L*(end-start)
            key_padding_mask = torch.nn.utils.rnn.pad_sequence(
                merge_key_padding_mask, batch_first=True,
                padding_value=True).squeeze(-1)

        out = self.attn(
            self._repeat(q, bs),  # Q * B * D
            k,  # L * B * D +  L * B * D
            v,
            key_padding_mask=key_padding_mask,
        )[0]
        #  out: Q * B * D
        x = out.permute(1, 0, 2)  # B * Q * D

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


404
def get_version_by_config(config: PretrainedConfig) -> tuple[int, ...]:
405
406
407
408
409
410
411
412
413
414
415
416
    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("."))


417
def _minicpmv_field_config(hf_inputs: Mapping[str, torch.Tensor]):
418
419
420
421
422
423
    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)

424
    return dict(
425
        pixel_values=MultiModalFieldConfig.batched("image"),
426
        image_sizes=MultiModalFieldConfig.batched("image"),
427
428
429
        tgt_sizes=MultiModalFieldConfig.batched("image"),
        image_embeds=MultiModalFieldConfig.batched("image"),
        video_pixel_values=MultiModalFieldConfig.batched("video"),
430
        video_image_sizes=MultiModalFieldConfig.batched("video"),
431
432
        video_tgt_sizes=MultiModalFieldConfig.batched("video"),
        video_embeds=MultiModalFieldConfig.batched("video"),
433
434
        image_token_id=MultiModalFieldConfig.shared("image", num_images),
        video_token_id=MultiModalFieldConfig.shared("video", num_videos),
435
436
437
438
439
440
441
442
    )


class MiniCPMVImageEmbeddingItems(DictEmbeddingItems):

    def __init__(
        self,
        data: Mapping[str, torch.Tensor],
443
444
445
446
        fields_factory: Callable[
            [Mapping[str, torch.Tensor]],
            Mapping[str, MultiModalFieldConfig],
        ],
447
448
449
450
451
    ) -> None:
        super().__init__(
            data,
            modality="image",
            required_fields={"image_embeds", "image_sizes"},
452
            fields_factory=fields_factory,
453
454
455
456
457
458
459
460
461
462
463
464
        )

    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],
465
466
467
468
        fields_factory: Callable[
            [Mapping[str, torch.Tensor]],
            Mapping[str, MultiModalFieldConfig],
        ],
469
470
471
472
473
    ) -> None:
        super().__init__(
            data,
            modality="video",
            required_fields={"video_embeds", "video_image_sizes"},
474
            fields_factory=fields_factory,
475
476
477
478
479
480
481
482
483
484
        )

    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"])


485
486
487
488
489
class MiniCPMVMultiModalDataParser(MultiModalDataParser):

    def _parse_image_data(
        self,
        data: Union[dict[str, torch.Tensor], ModalityData[ImageItem]],
490
    ) -> Optional[ModalityDataItems[Any, Any]]:
491
        if isinstance(data, dict):
492
493
            return MiniCPMVImageEmbeddingItems(
                data,
494
                fields_factory=_minicpmv_field_config,
495
496
            )

497
498
499
500
501
        return super()._parse_image_data(data)

    def _parse_video_data(
        self,
        data: Union[dict[str, torch.Tensor], ModalityData[VideoItem]],
502
    ) -> Optional[ModalityDataItems[Any, Any]]:
503
        if isinstance(data, dict):
504
505
            return MiniCPMVVideoEmbeddingItems(
                data,
506
                fields_factory=_minicpmv_field_config,
507
508
            )

509
510
511
512
513
514
515
516
517
518
        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()

519
520
    def get_hf_processor(self, **kwargs: object):
        hf_processor = self.ctx.get_hf_processor(**kwargs)
521
522
523
524
525
526
527
528
529

        # 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())

530
531
        return hf_processor

532
533
    def get_image_processor(self, **kwargs: object):
        return self.get_hf_processor(**kwargs).image_processor
534
535
536
537
538

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

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
539
        mm_limits = {"image": None}
tc-mb's avatar
tc-mb committed
540
        if self.get_model_version() in {(2, 6), (4, 0), (4, 5)}:
541
542
543
            mm_limits["video"] = None

        return mm_limits
544

545
546
547
548
549
550
551
552
553
554
    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()
555

556
557
        if version == (2, 0) or version == (2, 5):
            return image_processor.get_slice_image_placeholder(image_size)
558

559
560
561
562
563
564
        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,
        )
565

566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
    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,
        )

586
587
588
589
590
    def get_num_image_tokens(
        self,
        image_size: ImageSize,
        max_slice_nums: Optional[int] = None,
    ) -> int:
591
592
593
        image_processor = self.get_image_processor()

        grid = self.get_sliced_grid(
594
595
596
            image_size,
            max_slice_nums=max_slice_nums,
        )
597
598
599
600
        if grid is None:
            ncols = nrows = 0
        else:
            ncols, nrows = grid
601

602
        return (ncols * nrows + 1) * image_processor.image_feature_size
603
604
605

    def get_max_image_tokens(self) -> int:
        image_size = self.get_image_size_with_most_features()
606
607
608
609
        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)
610
611

    def get_image_size_with_most_features(self) -> ImageSize:
612
613
614
615
616
617
618
619
620
621
622
623
        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(),
        )

624
625
626
627
628
629
630
631
    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
632
633
634

    def get_video_max_slice_num(self) -> int:
        return 1
635

636
    def get_video_frame_size_with_most_features(self) -> ImageSize:
637
638
639
        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)
640

641
642
643
644
    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
645

646
647
648
649
650
651
652
    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)
653

654
        max_image_tokens = self.get_max_image_tokens() * max_images
655
656
        max_total_frames = self.get_max_video_frames(seq_len -
                                                     max_image_tokens)
657
658
        max_frames_per_video = min(max_total_frames // max(max_videos, 1),
                                   _MAX_FRAMES_PER_VIDEO)
659

660
        return max(max_frames_per_video, 1)
661
662


663
664
665
666
667
668
_I = TypeVar("_I",
             bound=MiniCPMVProcessingInfo,
             default=MiniCPMVProcessingInfo)


class MiniCPMVDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
669

670
671
672
673
674
675
676
677
678
679
    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(
680
681
682
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
683
        mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
684
    ) -> MultiModalDataDict:
685
686
687
688
689
690
691
692
        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 = \
693
            self.info.get_num_frames_with_most_features(seq_len, mm_counts)
694

695
696
697
        image_overrides = mm_options.get("image") if mm_options else None
        video_overrides = mm_options.get("video") if mm_options else None

698
        return {
699
700
701
            "image":
            self._get_dummy_images(width=image_width,
                                   height=image_height,
702
703
                                   num_images=num_images,
                                   overrides=image_overrides),
704
705
706
            "video": [
                self._get_dummy_images(width=video_width,
                                       height=video_height,
707
708
                                       num_images=num_video_frames,
                                       overrides=video_overrides)
709
710
711
712
            ] * num_videos,
        }


713
class MiniCPMVMultiModalProcessor(BaseMultiModalProcessor[_I]):
714
715
716
717
718
719
720

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

    def get_image_prompt_texts(self,
                               image_size: ImageSize,
                               image_idx: int = 0) -> str:
721
722
723
724
        return self.info.get_slice_image_placeholder(
            image_size,
            image_idx=image_idx,
        )
725
726
727

    def get_video_prompt_texts(self, image_size: ImageSize,
                               num_frames: int) -> str:
728
        return self.info.get_slice_image_placeholder(
729
730
731
732
733
            image_size=image_size,
            image_idx=0,
            max_slice_nums=self.info.get_video_max_slice_num(),
            use_image_id=False,
        ) * num_frames
734

735
736
737
738
    def process_images(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
739
        tok_kwargs: Mapping[str, object],
740
    ) -> Mapping[str, NestedTensors]:
741
742
743
744
745
        if (images := mm_data.get("images")) is None:
            return {}

        parsed_images = (self._get_data_parser().parse_mm_data({
            "image": images
746
747
        }).get_items("image",
                     (MiniCPMVImageEmbeddingItems, ImageProcessorItems)))
748

749
750
751
752
753
754
755
        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,
756
                tok_kwargs=tok_kwargs,
757
758
759
760
761
762
763
764
                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
765

766
767
768
769
    def process_videos(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
770
        tok_kwargs: Mapping[str, object],
771
    ) -> Mapping[str, NestedTensors]:
772
773
774
775
776
        if (videos := mm_data.get("videos")) is None:
            return {}

        parsed_videos = (self._get_data_parser().parse_mm_data({
            "video": videos
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
        }).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(),
                },
794
                tok_kwargs=tok_kwargs,
795
796
797
                out_keys={"pixel_values", "image_sizes", "tgt_sizes"},
            )

798
799
        video_inputs = {f"video_{k}": v for k, v in video_inputs.items()}

800
        tokenizer = self.info.get_tokenizer()
801
802
        unk_token_id = tokenizer.get_vocab()["<unk>"]
        video_inputs["video_token_id"] = torch.tensor(unk_token_id)
803

804
        return video_inputs
805

806
807
808
809
    def process_mm_inputs(
        self,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
810
        tok_kwargs: Mapping[str, object],
811
    ) -> Mapping[str, NestedTensors]:
812
        return {
813
814
            **self.process_images(mm_data, mm_kwargs, tok_kwargs),
            **self.process_videos(mm_data, mm_kwargs, tok_kwargs),
815
        }
816

817
    def _base_call_hf_processor(
818
        self,
819
820
        prompts: list[str],
        mm_data: Mapping[str, Sequence[object]],
821
        mm_kwargs: Mapping[str, object],
822
        tok_kwargs: Mapping[str, object],
823
824
        *,
        out_keys: set[str],
825
    ) -> dict[str, NestedTensors]:
826
        # This processor supports zipping prompt and mm_data together
tc-mb's avatar
tc-mb committed
827
        if self.info.get_model_version() in {(2, 6), (4, 0), (4, 5)}:
828
829
830
831
            inputs = super()._call_hf_processor(
                prompt=prompts,  # type: ignore
                mm_data=mm_data,
                mm_kwargs=mm_kwargs,
832
                tok_kwargs=tok_kwargs,
833
834
835
836
837
838
839
840
841
842
843
844
            )
        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,
845
                    tok_kwargs=tok_kwargs,
846
847
848
849
850
851
852
                )

                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}
853
854
855
856
857
858

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
859
        tok_kwargs: Mapping[str, object],
860
861
    ) -> BatchFeature:
        tokenizer = self.info.get_tokenizer()
862

863
864
        input_ids = torch.tensor([tokenizer.encode(prompt, **tok_kwargs)])
        mm_inputs = self.process_mm_inputs(mm_data, mm_kwargs, tok_kwargs)
865
866

        return BatchFeature({
867
            "input_ids": input_ids,
868
            **mm_inputs,
869
        })
870

871
    def _hf_processor_applies_updates(
872
873
874
875
        self,
        prompt_text: str,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
876
        tokenization_kwargs: Mapping[str, object],
877
878
879
    ) -> bool:
        return False

880
    def _get_prompt_updates(
881
882
883
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
884
        out_mm_kwargs: MultiModalKwargsItems,
885
    ) -> Sequence[PromptUpdate]:
tc-mb's avatar
tc-mb committed
886
887
888
889
890
891
892
893
894
895
896
897
        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
898

899
900
901
902
903
904
        def get_image_replacement(item_idx: int):
            images = mm_items.get_items(
                "image", (MiniCPMVImageEmbeddingItems, ImageProcessorItems))

            image_size = images.get_image_size(item_idx)

905
906
907
908
            return PromptUpdateDetails.select_text(
                self.get_image_prompt_texts(image_size, item_idx),
                "<unk>",
            )
909
910
911
912
913
914
915
916

        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)

917
918
919
920
            return PromptUpdateDetails.select_text(
                self.get_video_prompt_texts(frame_size, num_frames),
                "<unk>",
            )
921
922
923
924
925

        get_replacement = {
            "image": get_image_replacement,
            "video": get_video_replacement,
        }
926
927
928

        return [
            PromptReplacement(modality=modality,
tc-mb's avatar
tc-mb committed
929
                              target=pattern,
930
                              replacement=get_replacement[modality])
tc-mb's avatar
tc-mb committed
931
            for modality, pattern in placeholders
932
        ]
933

934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
    def _recompute_cached_prompt_update(
        self,
        cached_update: ResolvedPromptUpdate,
        new_item_idx: int,
    ) -> ResolvedPromptUpdate:
        new_update = super()._recompute_cached_prompt_update(
            cached_update,
            new_item_idx,
        )

        if cached_update.modality == "image":
            tokenizer = self.info.get_tokenizer()
            image_processor = self.info.get_image_processor()
            version = self.info.get_model_version()

            text = _seq2text(tokenizer, cached_update.content.full)
            prev_item_idx = cached_update.item_idx

            if version == (2, 0) or version == (2, 5):
                im_start = image_processor.im_start_token
                im_end = image_processor.im_end_token
            else:
                im_start = image_processor.im_id_start
                im_end = image_processor.im_id_end

            new_update = new_update.with_content(
                PromptUpdateDetails.select_text(
                    text.replace(
                        f"{im_start}{prev_item_idx}{im_end}",
                        f"{im_start}{new_item_idx}{im_end}",
                        1,
                    ),
                    "<unk>",
                ))

        return new_update

971
972
    def _get_mm_fields_config(
        self,
973
        hf_inputs: BatchFeature,
974
975
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
976
        return _minicpmv_field_config(hf_inputs)
977

978
979

class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
Jee Jee Li's avatar
Jee Jee Li committed
980
981
982
983
    """
    The abstract class of MiniCPMV can only be inherited, but cannot be
    instantiated.
    """
984

985
986
    supports_encoder_tp_data = True

987
988
989
990
991
992
993
994
995
    @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")

996
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
997
998
999
        config = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config
        quant_config = vllm_config.quant_config
1000
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
1001
        super().__init__()
1002
1003
1004
1005
        # 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
1006
1007
1008
        self.config = config
        self.multimodal_config = multimodal_config

1009
        self.version = get_version_by_config(self.config)
1010
1011
1012
1013
1014
        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
1015
1016
        self.vision_dim = (self.vpm.embed_dim if self.version == (2, 0) else
                           self.vpm.embeddings.embed_dim)
Alphi's avatar
Alphi committed
1017
        self.embed_dim = self.config.hidden_size
1018

1019
1020
1021
        self.resampler = self.init_resampler(self.embed_dim,
                                             self.vision_dim,
                                             quant_config=quant_config,
1022
1023
                                             prefix=maybe_prefix(
                                                 prefix, "resampler"))
1024

1025
        self.mm_token_ids = set[int]()
1026
1027
1028
        self.make_empty_intermediate_tensors = (
            self.llm.make_empty_intermediate_tensors)

1029
    def _parse_and_validate_vision_input(
Jee Jee Li's avatar
Jee Jee Li committed
1030
        self,
1031
        modality: str,
Jee Jee Li's avatar
Jee Jee Li committed
1032
        **kwargs: object,
1033
    ) -> Optional[MiniCPMVImageInputs]:
1034
1035
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)
1036

1037
        if pixel_values is None and image_embeds is None:
1038
1039
            return None

1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
        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)
1052

1053
            return MiniCPMVImageEmbeddingInputs(
1054
                type="image_embeds",
1055
                image_embeds=image_embeds_flat,
1056
            )
1057

1058
1059
1060
1061
        if not isinstance(pixel_values, (torch.Tensor, list)):
            raise ValueError(
                f"Incorrect type of pixel_values for {modality=}. "
                f"Got type: {type(pixel_values)}")
1062

1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
        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)
1073

1074
        return MiniCPMVImagePixelInputs(
1075
1076
            type="pixel_values",
            pixel_values=pixel_values_flat,
1077
1078
            tgt_sizes=tgt_sizes_flat,
            num_slices=num_slices_flat,
Jee Jee Li's avatar
Jee Jee Li committed
1079
        )
1080

1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
    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)

1117
1118
1119
1120
1121
        num_slices = image_input["num_slices"]
        return [
            e.flatten(0, 1)
            for e in image_features_flat.split(num_slices.tolist())
        ]
1122
1123
1124

    def _process_multimodal_inputs(self, modalities: dict):
        # The result multimodal_embeddings is tuple of tensors, with each
1125
        # tensor corresponding to a multimodal data item (image or video).
1126
1127
1128
1129
1130
1131
1132
1133
        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)
1134
                multimodal_embeddings += tuple(image_features)
1135
1136
1137
            if modality == "videos":
                video_input = modalities["videos"]
                video_features = self._process_vision_input(video_input)
1138
                multimodal_embeddings += tuple(video_features)
1139
1140
1141

        return multimodal_embeddings

1142
1143
1144
    def get_language_model(self) -> torch.nn.Module:
        return self.llm

1145
1146
    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
1147
1148
        modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
        if not modalities:
1149
            return []
1150
1151
1152

        return self._process_multimodal_inputs(modalities)

1153
1154
1155
1156
1157
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
1158
        inputs_embeds: Optional[torch.Tensor] = None,
Jee Jee Li's avatar
Jee Jee Li committed
1159
1160
        **kwargs: Any,
    ) -> torch.Tensor:
1161
        if intermediate_tensors is not None:
1162
1163
1164
            inputs_embeds = None

        hidden_states = self.llm.model(
1165
            input_ids=input_ids,
Jee Jee Li's avatar
Jee Jee Li committed
1166
1167
            positions=positions,
            intermediate_tensors=intermediate_tensors,
1168
            inputs_embeds=inputs_embeds,
Jee Jee Li's avatar
Jee Jee Li committed
1169
        )
1170
        return hidden_states
1171

1172
1173
1174
1175
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
1176
        return self.llm.compute_logits(hidden_states)
1177

1178
1179
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
1180
1181
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)
Jee Jee Li's avatar
Jee Jee Li committed
1182

1183
1184
1185
1186
1187
1188
1189
1190
    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
1191
1192
    def init_llm(
        self,
1193
        vllm_config: VllmConfig,
1194
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1195
1196
1197
    ) -> nn.Module:
        raise NotImplementedError

1198
1199
1200
1201
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1202
        prefix: str = "",
1203
    ) -> nn.Module:
Jee Jee Li's avatar
Jee Jee Li committed
1204
1205
        raise NotImplementedError

1206
1207
1208
1209
1210
    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
1211
1212
        raise NotImplementedError

1213
1214
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
Jee Jee Li's avatar
Jee Jee Li committed
1215
1216
1217
        raise NotImplementedError


1218
class MiniCPMV2_0(MiniCPMVBaseModel):
Jee Jee Li's avatar
Jee Jee Li committed
1219

1220
1221
    supports_encoder_tp_data = False

1222
1223
    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
1224
1225
1226
1227
        assert self.version == (2, 0)

    def init_llm(
        self,
1228
        vllm_config: VllmConfig,
1229
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1230
    ) -> nn.Module:
1231
        return MiniCPMForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1232

1233
1234
1235
1236
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1237
        prefix: str = "",
1238
    ) -> nn.Module:
1239
        # TODO: refactor vision model through timm wrapper from transformers
Jee Jee Li's avatar
Jee Jee Li committed
1240
1241
1242
1243
        try:
            import timm
        except ImportError:
            raise ImportError("Please install timm==0.9.10") from ImportError
1244

Jee Jee Li's avatar
Jee Jee Li committed
1245
1246
1247
1248
1249
1250
1251
1252
1253
        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,
            )

1254
1255
        model = model.to(dtype=torch.get_default_dtype())

Jee Jee Li's avatar
Jee Jee Li committed
1256
1257
1258
1259
1260
1261
1262
1263
1264
        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

1265
1266
1267
1268
1269
    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
1270
        with set_default_torch_dtype(torch.float16):
1271
1272
1273
1274
1275
1276
1277
1278
1279
            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
1280

1281
1282
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1283

1284
1285
1286
1287
1288
1289
1290
1291
1292
    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
1293
1294
        for pixel_value in pixel_values:
            H, W = pixel_value[0].shape[-2:]
1295
            tgt_size = (math.ceil(H / P_h), math.ceil(W / P_w))
Jee Jee Li's avatar
Jee Jee Li committed
1296
1297
1298
            vision_embedding = self.vpm.forward_features(
                pixel_value.unsqueeze(0).type(dtype))

1299
1300
1301
            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
1302

1303
        return torch.vstack(res)
Jee Jee Li's avatar
Jee Jee Li committed
1304
1305


1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
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
1318

1319
1320
    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
1321
1322
1323
1324
        assert self.version == (2, 5)

    def init_llm(
        self,
1325
        vllm_config: VllmConfig,
1326
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1327
    ) -> nn.Module:
1328
        return LlamaForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1329

1330
1331
1332
1333
    def init_vision_module(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig],
1334
        prefix: str = "",
1335
    ) -> nn.Module:
1336
1337
1338
1339
1340
1341
        model = Idefics2VisionTransformer(
            config.vision_config,
            quant_config=quant_config,
            prefix=prefix,
            use_data_parallel=self.use_data_parallel,
        )
Jee Jee Li's avatar
Jee Jee Li committed
1342
1343
1344
1345
        if self.config.drop_vision_last_layer:
            model.encoder.layers = model.encoder.layers[:-1]
        return model

1346
1347
1348
1349
1350
    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
1351
        with set_default_torch_dtype(torch.float16):
1352
1353
1354
1355
1356
1357
            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)
1358

1359
1360
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1361

1362
1363
1364
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]
Jee Jee Li's avatar
Jee Jee Li committed
1365
1366
        tgt_sizes = data["tgt_sizes"]

1367
1368
1369
1370
1371
        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
1372

1373
1374
1375
1376
1377
1378
        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
1379

1380
1381
1382
        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
1383

1384
        patch_attn_mask = torch.zeros((B, max_patches),
Jee Jee Li's avatar
Jee Jee Li committed
1385
1386
                                      dtype=torch.bool,
                                      device=device)
1387
1388
        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
1389

1390
1391
1392
1393
1394
1395
1396
        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
1397
1398


1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
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
1411

1412
1413
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
1414
        assert self.version == (2, 6)
Jee Jee Li's avatar
Jee Jee Li committed
1415
1416
1417

    def init_llm(
        self,
1418
        vllm_config: VllmConfig,
1419
        prefix: str = "",
Jee Jee Li's avatar
Jee Jee Li committed
1420
    ) -> nn.Module:
1421
        return Qwen2ForCausalLM(vllm_config=vllm_config, prefix=prefix)
Jee Jee Li's avatar
Jee Jee Li committed
1422

1423
1424
1425
    def init_vision_module(
        self,
        config: PretrainedConfig,
1426
        quant_config: Optional[QuantizationConfig] = None,
1427
        prefix: str = "",
1428
    ) -> nn.Module:
1429
1430
1431
1432
1433
1434
        model = Idefics2VisionTransformer(
            config.vision_config,
            quant_config=quant_config,
            prefix=prefix,
            use_data_parallel=self.use_data_parallel,
        )
Jee Jee Li's avatar
Jee Jee Li committed
1435
1436
1437
1438
        if self.config.drop_vision_last_layer:
            model.encoder.layers = model.encoder.layers[:-1]
        return model

1439
1440
1441
1442
1443
    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
1444
        with set_default_torch_dtype(torch.float16):
1445
            # The resampler in 2.6 remains consistent with the one in 2.5.
1446
1447
1448
1449
1450
1451
            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)
1452

1453
1454
        return resampler.to(device=current_platform.device_type,
                            dtype=torch.get_default_dtype())
Jee Jee Li's avatar
Jee Jee Li committed
1455

1456
1457
1458
    def get_vision_hidden_states(
            self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
        pixel_values = data["pixel_values"]
Jee Jee Li's avatar
Jee Jee Li committed
1459
1460
        tgt_sizes = data["tgt_sizes"]

1461
1462
1463
1464
1465
        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
1466

1467
1468
1469
1470
1471
1472
        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
1473

1474
1475
1476
        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
1477

1478
        patch_attn_mask = torch.zeros((B, max_patches),
Jee Jee Li's avatar
Jee Jee Li committed
1479
1480
                                      dtype=torch.bool,
                                      device=device)
1481
1482
1483
        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
1484
        vision_embedding = self.vpm(
1485
1486
            all_pixel_values,
            patch_attention_mask=patch_attn_mask.unsqueeze(1),
Jee Jee Li's avatar
Jee Jee Li committed
1487
            tgt_sizes=tgt_sizes,
1488
        )
Jee Jee Li's avatar
Jee Jee Li committed
1489
1490
1491

        return self.resampler(vision_embedding, tgt_sizes)

tc-mb's avatar
tc-mb committed
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
    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)
1535
1536
1537
1538
1539
1540
        model = Idefics2VisionTransformer(
            config.vision_config,
            quant_config=quant_config,
            prefix=prefix,
            use_data_parallel=self.use_data_parallel,
        )
tc-mb's avatar
tc-mb committed
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
        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
1607

tc-mb's avatar
tc-mb committed
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
class MiniCPMV4_5(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, 5)

    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 Qwen3ForCausalLM(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)
1644
1645
1646
1647
1648
1649
        model = Idefics2VisionTransformer(
            config.vision_config,
            quant_config=quant_config,
            prefix=prefix,
            use_data_parallel=self.use_data_parallel,
        )
tc-mb's avatar
tc-mb committed
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
        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 = Resampler4_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"]
        temporal_ids = data.get('temporal_ids', None)

        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)
        all_temporal_ids = None if temporal_ids is None else flatten_2d_lists(
            temporal_ids)
        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, all_temporal_ids)

    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)


1720
1721
1722
_SUPPORT_VERSION = {
    (2, 0): MiniCPMV2_0,
    (2, 5): MiniCPMV2_5,
1723
    (2, 6): MiniCPMV2_6,
tc-mb's avatar
tc-mb committed
1724
    (4, 0): MiniCPMV4_0,
tc-mb's avatar
tc-mb committed
1725
    (4, 5): MiniCPMV4_5,
1726
1727
1728
}


1729
1730
1731
1732
1733
@MULTIMODAL_REGISTRY.register_processor(
    MiniCPMVMultiModalProcessor,
    info=MiniCPMVProcessingInfo,
    dummy_inputs=MiniCPMVDummyInputsBuilder)
class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
Jee Jee Li's avatar
Jee Jee Li committed
1734
1735
1736
1737
1738
    """
    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.
    """
1739

1740
    def __new__(cls, *, vllm_config: VllmConfig, prefix: str = ""):
1741
        config = vllm_config.model_config.hf_config
Jee Jee Li's avatar
Jee Jee Li committed
1742
1743
1744
1745
1746
1747
1748
1749
1750
        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
1751
1752
        instance_cls = _SUPPORT_VERSION.get(version)
        if instance_cls is None:
tc-mb's avatar
tc-mb committed
1753
1754
1755
1756
            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}")
1757
1758
1759
1760
1761
1762
1763

        # 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)