gemma3_mm.py 25.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
import math
4
from collections.abc import Iterable, Mapping, Sequence
5
from typing import Annotated, Any, Literal, Optional
6
7
8

import torch
from torch import nn
9
10
from transformers import BatchFeature, Gemma3Config, Gemma3Processor
from transformers.models.gemma3.processing_gemma3 import Gemma3ProcessorKwargs
11

12
import vllm.envs as envs
13
14
15
from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
16
from vllm.model_executor.models.module_mapping import MultiModelKeys
17
18
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
19
                                    MultiModalKwargsItems)
20
21
from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                   MultiModalDataItems)
22
# yapf: disable
23
from vllm.multimodal.processing import (BaseMultiModalProcessor,
24
25
26
                                        BaseProcessingInfo,
                                        MultiModalPromptUpdates,
                                        MultiModalPromptUpdatesApplyResult,
27
                                        PlaceholderFeaturesInfo,
28
29
                                        PromptReplacement, PromptUpdate,
                                        PromptUpdateDetails,
30
31
                                        replace_token_matches)
# yapf: enable
32
from vllm.multimodal.profiling import BaseDummyInputsBuilder
33
from vllm.sequence import IntermediateTensors
34
from vllm.utils.tensor_schema import TensorSchema, TensorShape
35

36
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
37
                         SupportsMultiModal, SupportsPP)
38
from .siglip import SiglipVisionModel
39
40
41
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
                    init_vllm_registered_model, maybe_prefix,
                    merge_multimodal_embeddings)
42
43
44
45

logger = init_logger(__name__)


46
class Gemma3ImagePixelInputs(TensorSchema):
47
    """
48
49
50
51
52
53
54
    Dimensions:
        - p: Number of patches total (over each image over each prompt in the
          batch)
        - c: Number of channels (3)
        - h: Height of each patch
        - w: Width of each patch
        - bn: Batch size * number of images
55
    """
56
57
58
    type: Literal["pixel_values"] = "pixel_values"

    pixel_values: Annotated[torch.Tensor, TensorShape("p", 3, "h", "w")]
59

60
    num_patches: Annotated[torch.Tensor, TensorShape("bn")]
61

62
63
64
65
66
67

Gemma3ImageInputs = Gemma3ImagePixelInputs


class Gemma3ProcessingInfo(BaseProcessingInfo):

68
69
70
    def get_hf_config(self):
        return self.ctx.get_hf_config(Gemma3Config)

71
72
73
    def get_hf_processor(self, **kwargs: object):
        return self.ctx.get_hf_processor(Gemma3Processor, **kwargs)

74
75
76
    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": None}

77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
    def _resolve_image_kwargs(
        self,
        processor: Gemma3Processor,
        keys: set[str],
    ) -> dict[str, Any]:
        image_processor = processor.image_processor
        kwargs = processor._merge_kwargs(
            Gemma3ProcessorKwargs,
            tokenizer_init_kwargs=processor.tokenizer.init_kwargs,
        )

        images_kwargs = kwargs["images_kwargs"]

        def _resolve_kw(key: str):
            val = getattr(image_processor, key)
            if val is None:
                val = images_kwargs[key]

            return val

        return {k: _resolve_kw(k) for k in keys}

    def get_num_crops(
        self,
        *,
        image_width: int,
        image_height: int,
        processor: Optional[Gemma3Processor],
    ) -> int:
        if processor is None:
            processor = self.get_hf_processor()

        images_kwargs = self._resolve_image_kwargs(
            processor, {
                "do_pan_and_scan", "pan_and_scan_min_crop_size",
                "pan_and_scan_max_num_crops",
                "pan_and_scan_min_ratio_to_activate"
            })

        do_pan_and_scan = images_kwargs["do_pan_and_scan"]
        pan_and_scan_min_crop_size = images_kwargs[
            "pan_and_scan_min_crop_size"]
        pan_and_scan_max_num_crops = images_kwargs[
            "pan_and_scan_max_num_crops"]
        pan_and_scan_min_ratio_to_activate = images_kwargs[
            "pan_and_scan_min_ratio_to_activate"]

        if not do_pan_and_scan:
            return 0

127
128
129
130
131
        if envs.VLLM_USE_V1:
            logger.warning_once(
                "`do_pan_and_scan=True` has suboptimal results on V1 "
                "because of the simplified attention pattern being used.")

132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
        # Based on Gemma3ImageProcessor.pan_and_scan
        if image_width >= image_height:
            if image_width / image_height < pan_and_scan_min_ratio_to_activate:
                return 0

            num_crops_w = min(
                int(math.floor(image_width / pan_and_scan_min_crop_size)),
                int(math.floor(image_width / image_height + 0.5)),
            )

            num_crops_w = max(2, num_crops_w)
            num_crops_w = min(pan_and_scan_max_num_crops, num_crops_w)
            num_crops_h = 1
        else:
            if image_height / image_width < pan_and_scan_min_ratio_to_activate:
                return 0

            num_crops_h = min(
                int(math.floor(image_height / pan_and_scan_min_crop_size)),
                int(math.floor(image_height / image_width + 0.5)),
            )

            num_crops_h = max(2, num_crops_h)
            num_crops_h = min(pan_and_scan_max_num_crops, num_crops_h)
            num_crops_w = 1

        crop_size_w = int(math.ceil(image_width / num_crops_w))
        crop_size_h = int(math.ceil(image_height / num_crops_h))

        if min(crop_size_w, crop_size_h) < pan_and_scan_min_crop_size:
            return 0

        return num_crops_w * num_crops_h

    def get_image_repl(
        self,
        *,
        image_width: int,
        image_height: int,
        processor: Optional[Gemma3Processor],
172
    ) -> PromptUpdateDetails[str]:
173
174
175
        if processor is None:
            processor = self.get_hf_processor()

176
        boi_token = processor.boi_token
177
178
179
180
181
182
183
184

        num_crops = self.get_num_crops(
            image_width=image_width,
            image_height=image_height,
            processor=processor,
        )

        if num_crops == 0:
185
            image_text = boi_token
186
        else:
187
            crops_image_tokens = " ".join(boi_token for _ in range(num_crops))
188
            image_text = (
189
                f"Here is the original image {boi_token} and here are some "
190
191
                f"crops to help you see better {crops_image_tokens}")

192
        repl_full = image_text.replace(boi_token,
193
194
                                       processor.full_image_sequence)

195
196
197
198
199
        tokenizer = processor.tokenizer
        vocab = tokenizer.get_vocab()
        image_token_id = vocab[tokenizer.image_token]

        return PromptUpdateDetails.select_token_id(repl_full, image_token_id)
200
201
202
203
204
205

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
206
        processor: Optional[Gemma3Processor],
207
    ) -> int:
208
209
210
211
        if processor is None:
            processor = self.get_hf_processor()

        num_crops = self.get_num_crops(
212
213
214
215
            image_width=image_width,
            image_height=image_height,
            processor=processor,
        )
216
        image_seq_len = processor.image_seq_length
217

218
        return (num_crops + 1) * image_seq_len
219
220

    def get_image_size_with_most_features(self) -> ImageSize:
221
222
223
224
225
226
227
228
229
        processor = self.get_hf_processor()

        images_kwargs = self._resolve_image_kwargs(
            processor, {"pan_and_scan_max_num_crops"})
        max_num_crops = images_kwargs["pan_and_scan_max_num_crops"]

        # Result in the max possible feature size (h:w = max_num_crops:1)
        return ImageSize(height=50 * max_num_crops, width=50)

230
231
232

class Gemma3DummyInputsBuilder(BaseDummyInputsBuilder[Gemma3ProcessingInfo]):

233
234
235
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

236
237
        processor = self.info.get_hf_processor()
        image_token = processor.boi_token
238

239
240
241
242
243
244
245
        return image_token * num_images

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> MultiModalDataDict:
246
        num_images = mm_counts.get("image", 0)
247

248
249
250
        target_width, target_height = \
            self.info.get_image_size_with_most_features()

251
        return {
252
253
254
255
256
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images)
        }
257

258
259
260
261
262
263
264
265

class Gemma3MultiModalProcessor(BaseMultiModalProcessor[Gemma3ProcessingInfo]):

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
266
        tok_kwargs: Mapping[str, object],
267
    ) -> BatchFeature:
268
269
270
271
        processed_outputs = super()._call_hf_processor(
            prompt,
            mm_data,
            mm_kwargs,
272
            tok_kwargs,
273
274
        )

275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
        # HF processor pops the `num_crops` kwarg, which is needed by vLLM
        if (images := mm_data.get("images")) is not None:
            parsed_images = (self._get_data_parser().parse_mm_data({
                "image":
                images
            }).get_items("image", ImageProcessorItems))
            image_sizes = [
                parsed_images.get_image_size(i)
                for i in range(len(parsed_images))
            ]
            hf_processor = self.info.get_hf_processor(**mm_kwargs)

            num_crops = [
                self.info.get_num_crops(image_width=size.width,
                                        image_height=size.height,
                                        processor=hf_processor)
                for size in image_sizes
            ]
            processed_outputs["num_crops"] = torch.tensor(num_crops)

        return processed_outputs

297
298
299
300
301
    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
302
303
304
305
306
307
308
        num_crops = hf_inputs.get("num_crops", torch.empty(0))

        return dict(
            pixel_values=MultiModalFieldConfig.flat_from_sizes(
                "image", num_crops + 1),
            num_crops=MultiModalFieldConfig.batched("image"),
        )
309
310
311
312
313

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, Any],
314
        out_mm_kwargs: MultiModalKwargsItems,
315
316
    ) -> Sequence[PromptUpdate]:
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
317
        image_token = hf_processor.boi_token
318
319

        def get_replacement_gemma3(item_idx: int):
320
321
322
323
324
325
326
            images = mm_items.get_items("image", ImageProcessorItems)

            image_size = images.get_image_size(item_idx)
            return self.info.get_image_repl(
                image_width=image_size.width,
                image_height=image_size.height,
                processor=hf_processor,
327
328
329
330
331
            )

        return [
            PromptReplacement(
                modality="image",
332
                target=image_token,
333
334
335
336
                replacement=get_replacement_gemma3,
            )
        ]

337
338
339
    def _apply_token_matches(
        self,
        prompt: list[int],
340
341
342
343
        mm_prompt_updates: MultiModalPromptUpdates,
    ) -> tuple[list[int], MultiModalPromptUpdatesApplyResult]:
        token_ids, res = super()._apply_token_matches(prompt,
                                                      mm_prompt_updates)
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

        # "\n\n\n" and "\n\n\n\n" are single tokens
        # Since our replacement can insert "\n\n" next to "\n"
        # tokens, we have to combine them to be consistent with
        # the output of the tokenizer
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()
        newline_1 = vocab["\n"]
        newline_2 = vocab["\n\n"]
        newline_3 = vocab["\n\n\n"]
        newline_4 = vocab["\n\n\n\n"]

        token_ids = replace_token_matches(
            token_ids,
            [newline_1, newline_2],
            [newline_3],
        )
        token_ids = replace_token_matches(
            token_ids,
            [newline_2, newline_1],
            [newline_3],
        )
        token_ids = replace_token_matches(
            token_ids,
            [newline_2, newline_2],
            [newline_4],
        )

372
        return token_ids, res
373
374
375
376

    def _find_mm_placeholders(
        self,
        new_token_ids: list[int],
377
        mm_prompt_updates: MultiModalPromptUpdates,
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
    ) -> Mapping[str, list[PlaceholderFeaturesInfo]]:
        # We need to detect "\n\n" inside "\n\n\n" and "\n\n\n\n"
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()
        newline_1 = vocab["\n"]
        newline_2 = vocab["\n\n"]
        newline_3 = vocab["\n\n\n"]
        newline_4 = vocab["\n\n\n\n"]

        def get_repl_toks(tok: int) -> list[int]:
            if tok == newline_3:
                return [newline_1, newline_2]
            if tok == newline_4:
                return [newline_2, newline_2]

            return [tok]

        repl_token_ids = list[int]()
        repl_orig_idxs = list[int]()
        for orig_idx, orig_tok in enumerate(new_token_ids):
            repl_toks = get_repl_toks(orig_tok)
            repl_token_ids.extend(repl_toks)
            repl_orig_idxs.extend(orig_idx for _ in range(len(repl_toks)))

402
403
        repls = super()._find_mm_placeholders(repl_token_ids,
                                              mm_prompt_updates)
404
405
406
407
408
409
410
411

        return {
            modality: [
                PlaceholderFeaturesInfo(
                    modality=p.modality,
                    item_idx=p.item_idx,
                    start_idx=repl_orig_idxs[p.start_idx],
                    tokens=p.tokens,
412
                    is_embed=p.is_embed,
413
414
415
416
417
                ) for p in placeholders
            ]
            for modality, placeholders in repls.items()
        }

418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461

class Gemma3MultiModalProjector(nn.Module):

    def __init__(self, config: Gemma3Config):
        super().__init__()

        self.mm_input_projection_weight = nn.Parameter(
            torch.zeros(config.vision_config.hidden_size,
                        config.text_config.hidden_size))

        self.mm_soft_emb_norm = GemmaRMSNorm(
            config.vision_config.hidden_size,
            eps=config.vision_config.layer_norm_eps)

        self.patches_per_image = int(config.vision_config.image_size //
                                     config.vision_config.patch_size)
        self.tokens_per_side = int(config.mm_tokens_per_image**0.5)
        self.kernel_size = self.patches_per_image // self.tokens_per_side
        self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size,
                                     stride=self.kernel_size)

    def forward(self, vision_outputs: torch.Tensor):
        batch_size, _, seq_length = vision_outputs.shape

        reshaped_vision_outputs = vision_outputs.transpose(1, 2)
        reshaped_vision_outputs = reshaped_vision_outputs.reshape(
            batch_size, seq_length, self.patches_per_image,
            self.patches_per_image)
        reshaped_vision_outputs = reshaped_vision_outputs.contiguous()

        pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
        pooled_vision_outputs = pooled_vision_outputs.flatten(2)
        pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2)

        normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)

        projected_vision_outputs = torch.matmul(
            normed_vision_outputs, self.mm_input_projection_weight)
        return projected_vision_outputs.type_as(vision_outputs)


@MULTIMODAL_REGISTRY.register_processor(Gemma3MultiModalProcessor,
                                        info=Gemma3ProcessingInfo,
                                        dummy_inputs=Gemma3DummyInputsBuilder)
462
class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
463
                                     SupportsLoRA):
464
465
466
467
468
469
470
471
472
473
474
475
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

476
477
478
479
480
481
482
483
484
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            # mapping for new names in checkpoint saved after transformers v4.52
            "model.language_model.": "language_model.model.",
            "model.vision_tower.": "vision_tower.",
            "model.multi_modal_projector.": "multi_modal_projector.",
            "lm_head.": "language_model.lm_head.",
        })

485
486
487
488
489
490
491
    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<start_of_image>"

        raise ValueError("Only image modality is supported")

492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.quant_config = quant_config
        self.multimodal_config = multimodal_config

        self.vision_tower = SiglipVisionModel(config.vision_config,
                                              quant_config,
                                              prefix=maybe_prefix(
                                                  prefix, "vision_tower"))
        self.multi_modal_projector = Gemma3MultiModalProjector(config)

        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=["Gemma3ForCausalLM"],
        )
        logit_scale = getattr(config, "logit_scale", 1.0)
514
515
516
517
518

        if hasattr(self.language_model, "logits_processor"):
            # The logits processor can be unset if we're using
            # automatic conversion to pooling model.
            self.language_model.logits_processor.scale *= logit_scale
519
520
521
522

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

523
524
525
526
    @property
    def dtype(self):
        return next(self.parameters()).dtype

527
528
529
    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[Gemma3ImageInputs]:
        pixel_values = kwargs.pop("pixel_values", None)
530
        num_crops = kwargs.pop("num_crops", None)
531
532
533
534
535
        image_embeds = kwargs.pop("image_embeds", None)
        assert image_embeds is None, "Gemma3 does not support image_embeds."
        if pixel_values is None:
            return None

536
        if not isinstance(pixel_values, (torch.Tensor, list)):
537
538
539
            raise ValueError("Incorrect type of pixel values. "
                             f"Got type: {type(pixel_values)}")

540
        if not isinstance(num_crops, (torch.Tensor, list)):
541
            raise ValueError("Incorrect type of num_crops. "
542
543
                             f"Got type: {type(num_crops)}")

544
        image_size = self.config.vision_config.image_size
545

546
        return Gemma3ImagePixelInputs(
547
548
549
550
551
552
            pixel_values=flatten_bn(pixel_values, concat=True),
            num_patches=flatten_bn(num_crops, concat=True) + 1,
            resolve_bindings={
                "h": image_size,
                "w": image_size
            })
553
554
555
556
557
558

    def _image_pixels_to_features(
        self,
        vision_tower: SiglipVisionModel,
        pixel_values: torch.Tensor,
    ) -> torch.Tensor:
559
        return vision_tower(pixel_values)
560
561
562
563

    def _process_image_input(
        self,
        image_input: Gemma3ImageInputs,
564
    ) -> list[torch.Tensor]:
565
        assert self.vision_tower is not None
566
567

        pixel_values = image_input["pixel_values"]
568
569
570
        num_patches = image_input["num_patches"]

        image_features = self._image_pixels_to_features(
571
572
573
            self.vision_tower,
            pixel_values,
        )
574
575
        image_embeds = self.multi_modal_projector(image_features)

576
577
578
        return [
            e.flatten(0, 1) for e in image_embeds.split(num_patches.tolist())
        ]
579

580
581
582
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

583
584
    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
585
586
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
587
            return []
588

589
        return self._process_image_input(image_input)
590
591
592
593

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
594
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
595
    ) -> torch.Tensor:
596
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
597
598
        if multimodal_embeddings is not None \
            and len(multimodal_embeddings) != 0:
599
            inputs_embeds = merge_multimodal_embeddings(
600
601
                input_ids,
                inputs_embeds,
602
                multimodal_embeddings,
603
604
                self.config.image_token_index,
            )
605
606
607
608
609
610
611
        return inputs_embeds

    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
                intermediate_tensors: Optional[IntermediateTensors] = None,
                inputs_embeds: Optional[torch.Tensor] = None,
612
                **kwargs: object) -> IntermediateTensors:
613
614
615
616
617
618
619
620
621
622
        if intermediate_tensors is not None:
            inputs_embeds = None

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility.
        elif inputs_embeds is None:
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)

            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
623
            if (vision_embeddings is not None) and len(vision_embeddings) != 0:
624
625
626
                kwargs = self.prepare_attn_masks(
                    input_ids,
                    positions,
627
628
629
                    mask_dtype=self.dtype,
                    **kwargs,
                )
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
            input_ids = None

        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds,
                                                  **kwargs)

        return hidden_states

    def prepare_attn_masks(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        mask_dtype: torch.dtype,
        **kwargs,
    ):
        kwargs["has_images"] = True
        # NOTE(woosuk): Here, we distinguish the sequences by the position id 0.
        # This is a HACK. Fix this.
650
651
        start_indices = (positions == 0).cpu().nonzero()
        num_seqs = len(start_indices)
652
653
        seq_lens = []
        for i in range(num_seqs):
654
            start_idx = start_indices[i].item()
655
            if i < num_seqs - 1:
656
                end_idx = start_indices[i + 1].item()
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
            else:
                end_idx = len(input_ids)
            seq_lens.append(end_idx - start_idx)
        kwargs["seq_lens"] = seq_lens

        global_attn_masks = []
        local_attn_masks = []
        start_idx = 0
        for seq_len in seq_lens:
            end_idx = start_idx + seq_len
            input_token_ids = input_ids[start_idx:end_idx]
            start_idx = end_idx
            # Create a global causal mask.
            global_attn_mask = torch.empty(
                1,
                1,
                seq_len,
                seq_len,
                dtype=mask_dtype,
                device=input_ids.device,
            )
            global_attn_mask.fill_(float("-inf"))
            # Fill the lower triangle with 0.
            global_attn_mask = global_attn_mask.triu(diagonal=1)

            # Consider the bidirectional attention between image tokens.
            img_mask = torch.zeros_like(global_attn_mask)
            img_pos = (input_token_ids == self.config.image_token_index)
            img_mask[:, :, :, img_pos] += 1
            img_mask[:, :, img_pos, :] += 1
            global_attn_mask = torch.where(img_mask == 2, 0, global_attn_mask)
            global_attn_masks.append(global_attn_mask)

690
691
            sliding_window = self.config.text_config.sliding_window
            if sliding_window is not None:
692
693
694
                # Create a local causal mask with sliding window (1024).
                local_attn_mask = torch.ones_like(global_attn_mask)
                local_attn_mask = torch.tril(local_attn_mask,
695
                                             diagonal=-sliding_window)
696
697
698
                local_attn_mask = torch.where(local_attn_mask == 0,
                                              global_attn_mask, float("-inf"))
                local_attn_masks.append(local_attn_mask)
699
700
701
702
703
704
705
706
        kwargs["global_attn_masks"] = global_attn_masks
        kwargs["local_attn_masks"] = local_attn_masks
        return kwargs

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> Optional[torch.Tensor]:
707
        return self.language_model.compute_logits(hidden_states)
708

709
710
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
711
        loader = AutoWeightsLoader(self)
712
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
713
714
715
716
717
718
719
720
721

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="language_model",
            connector="multi_modal_projector",
            tower_model="vision_tower")