pixtral.py 47.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import math
5
from collections.abc import Iterable, Mapping, Sequence
Patrick von Platen's avatar
Patrick von Platen committed
6
from dataclasses import dataclass, fields
7
from functools import cached_property
8
from typing import Literal, Optional, TypedDict, Union
Patrick von Platen's avatar
Patrick von Platen committed
9
10
11
12

import torch
import torch.nn as nn
import torch.nn.functional as F
13
14
15
from mistral_common.protocol.instruct.messages import (ImageChunk, TextChunk,
                                                       UserMessage)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
16
from mistral_common.tokens.tokenizers.multimodal import ImageEncoder
Patrick von Platen's avatar
Patrick von Platen committed
17
from PIL import Image
18
19
from transformers import PixtralVisionConfig, TensorType
from transformers.image_utils import ImageInput
20
from transformers.models.pixtral.image_processing_pixtral import (
21
    _num_image_tokens as _get_pixtral_hf_num_image_tokens)
22
from transformers.models.pixtral.modeling_pixtral import (
23
    PixtralRotaryEmbedding, apply_rotary_pos_emb, position_ids_in_meshgrid)
24
from transformers.tokenization_utils_base import TextInput
Patrick von Platen's avatar
Patrick von Platen committed
25

26
from vllm.config import VllmConfig
27
from vllm.distributed import divide, get_tensor_model_parallel_world_size
28
from vllm.model_executor.layers.activation import get_act_and_mul_fn
Patrick von Platen's avatar
Patrick von Platen committed
29
from vllm.model_executor.layers.layernorm import RMSNorm
30
31
32
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
Patrick von Platen's avatar
Patrick von Platen committed
33
34
35
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
36
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
37
38
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                    NestedTensors)
39
40
41
from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                   MultiModalDataItems)
from vllm.multimodal.processing import (BaseMultiModalProcessor,
42
43
44
                                        BaseProcessingInfo, MultiModalHashes,
                                        PromptReplacement, PromptUpdate,
                                        PromptUpdateDetails)
45
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
46
47
48
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.tokenizer import (MistralTokenizer,
                                               cached_tokenizer_from_config)
Patrick von Platen's avatar
Patrick von Platen committed
49

50
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
51
from .utils import (flatten_bn, init_vllm_registered_model, maybe_prefix,
52
                    merge_multimodal_embeddings)
53
from .vision import VisionEncoderInfo, resolve_visual_encoder_outputs
Patrick von Platen's avatar
Patrick von Platen committed
54

55
56
57
58
59
60
try:
    from xformers import ops as xops
    USE_XFORMERS_OPS = True
except ImportError:
    USE_XFORMERS_OPS = False

Patrick von Platen's avatar
Patrick von Platen committed
61
62
PATCH_MERGE = "patch_merge"

Patrick von Platen's avatar
Patrick von Platen committed
63

64
65
class PixtralImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
Patrick von Platen's avatar
Patrick von Platen committed
66

67
68
69
    images: Union[torch.Tensor, list[torch.Tensor]]
    """
    Shape: `(batch_size * num_images, num_channels, image_width, image_height)`
70

71
    The result of stacking `ImageEncoding.tokens` from each prompt.
72
    """
Patrick von Platen's avatar
Patrick von Platen committed
73
74


75
76
77
class PixtralProcessorAdapter:
    """
    Provide a HF-compatible interface for
78
    `mistral_common.tokens.tokenizers.multimodal.ImageEncoder`.
79
    """
Patrick von Platen's avatar
Patrick von Platen committed
80

81
82
    def __init__(self, tokenizer: MistralTokenizer) -> None:
        super().__init__()
Patrick von Platen's avatar
Patrick von Platen committed
83

84
        self.tokenizer = tokenizer
Patrick von Platen's avatar
Patrick von Platen committed
85

86
87
88
89
90
    @property
    def image_processor(self) -> ImageEncoder:
        image_encoder = self.tokenizer.instruct.mm_encoder
        assert isinstance(image_encoder, ImageEncoder)
        return image_encoder
91

92
93
94
    @cached_property
    def image_break_id(self) -> int:
        return self.image_processor.special_ids.img_break
Patrick von Platen's avatar
Patrick von Platen committed
95

96
97
98
    @cached_property
    def image_token_id(self) -> int:
        return self.image_processor.special_ids.img
Patrick von Platen's avatar
Patrick von Platen committed
99

100
101
102
    @cached_property
    def image_end_id(self) -> int:
        return self.image_processor.special_ids.img_end
Patrick von Platen's avatar
Patrick von Platen committed
103

104
105
106
    @cached_property
    def image_size(self) -> int:
        return self.image_processor.mm_config.max_image_size
Patrick von Platen's avatar
Patrick von Platen committed
107

108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
    @cached_property
    def patch_size(self) -> int:
        return self.image_processor.mm_config.image_patch_size

    def __call__(
        self,
        text: Optional[Union[TextInput, list[TextInput]]] = None,
        images: Optional[Union[ImageInput, list[ImageInput]]] = None,
        return_tensors: Optional[Union[str, TensorType]] = None,
        **kwargs,
    ) -> Mapping[str, NestedTensors]:
        if text is None:
            text = []
        if not isinstance(text, list):
            text = [text]
        if images is None:
            images = []
        if not isinstance(images, list):
            images = [images]

        if not images:
            input_ids = self.tokenizer(text).input_ids

            return {"input_ids": torch.tensor(input_ids)}

        # Allow dummy text, which is used for profiling as well as token inputs
        if any(len(t) > 0 for t in text):
            raise ValueError(
                "You've passed text inputs instead of token inputs. "
                "Make sure to process your input via `mistral_common`'s "
                "tokenizer or pass a chat completion request. "
                "For more info, see: "
                "https://github.com/vllm-project/vllm/issues/8411.")

        images_processed = list[torch.Tensor]()
        images_tokens = list[torch.Tensor]()

        for image in images:
            image_inputs = self.image_processor(ImageChunk(image=image))
            image_processed = torch.tensor(image_inputs.image)
            image_tokens = torch.tensor(image_inputs.tokens)

            images_processed.append(image_processed)
            images_tokens.append(image_tokens)

        return {
            "input_ids": torch.cat(images_tokens)[None].expand(len(text), -1),
            "images": images_processed,
        }


class PixtralProcessingInfo(BaseProcessingInfo):

    def get_tokenizer(self) -> MistralTokenizer:
        tokenizer = cached_tokenizer_from_config(self.ctx.model_config)
        if not isinstance(tokenizer, MistralTokenizer):
            raise ValueError("This model requires `--tokenizer-mode mistral`")

        return tokenizer

    def get_hf_processor(self) -> PixtralProcessorAdapter:
        return PixtralProcessorAdapter(self.get_tokenizer())

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

    def get_vision_config(
        self,
        processor: Optional[PixtralProcessorAdapter] = None,
    ):
        if processor is None:
            processor = self.get_hf_processor()

        return PixtralVisionConfig(
            image_size=processor.image_size,
            patch_size=processor.patch_size,
        )

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

        ncols, nrows = processor.image_processor._image_to_num_tokens(
            Image.new("RGB", (image_width, image_height)))

199
        return ncols * nrows
200
201
202
203
204
205
206
207
208
209

    def get_image_size_with_most_features(self) -> ImageSize:
        image_processor = self.get_hf_processor().image_processor
        max_image_size = image_processor.mm_config.max_image_size

        return ImageSize(width=max_image_size, height=max_image_size)


class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):

210
211
212
213
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        return ""

    def get_dummy_mm_data(
214
215
216
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
217
    ) -> MultiModalDataDict:
218
219
220
221
222
        num_images = mm_counts.get("image", 0)

        target_width, target_height = \
            self.info.get_image_size_with_most_features()

223
        return {
224
225
226
227
228
229
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images)
        }

230
231
232
233
234
235
236
237
238
239
    def get_dummy_processor_inputs(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> ProcessorInputs:
        tokenizer = self.info.get_tokenizer()

        dummy_text = self.get_dummy_text(mm_counts)
        dummy_mm_data = self.get_dummy_mm_data(seq_len, mm_counts)
        dummy_images = dummy_mm_data.get("image", [])
240
        tokenization_kwargs = {"truncation": False}
241
242
243
244
245
246
247
248
249
250

        request = ChatCompletionRequest(messages=[
            UserMessage(content=[
                TextChunk(text=dummy_text),
                *(ImageChunk(image=image) for image in dummy_images),
            ]),
        ])
        res = tokenizer.mistral.encode_chat_completion(request)
        dummy_tokens = res.tokens

251
252
253
        return ProcessorInputs(prompt=dummy_tokens,
                               mm_data=dummy_mm_data,
                               tokenization_kwargs=tokenization_kwargs)
254

255
256
257

class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo]
                                 ):
Patrick von Platen's avatar
Patrick von Platen committed
258

259
260
261
262
263
    def _get_mm_fields_config(
        self,
        hf_inputs: Mapping[str, NestedTensors],
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
264
        return dict(images=MultiModalFieldConfig.batched("image"))
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> Sequence[PromptUpdate]:
        processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)

        image_break_id = processor.image_break_id
        image_token_id = processor.image_token_id
        image_end_id = processor.image_end_id

        def get_replacement(item_idx: int):
            images = mm_items.get_items("image", ImageProcessorItems)
            image_size = images.get_image_size(item_idx)

            ncols, nrows = processor.image_processor._image_to_num_tokens(
                Image.new("RGB", (image_size.width, image_size.height)))

            tokens = ([image_token_id] * ncols + [image_break_id]) * nrows
            tokens[-1] = image_end_id

288
            return PromptUpdateDetails.select_token_id(tokens, image_token_id)
289
290
291
292
293
294
295
296
297
298
299
300
301
302

        return [
            PromptReplacement(
                modality="image",
                target="",  # Never match the prompt (see below note)
                replacement=get_replacement,
            ),
        ]

    def _cached_apply_hf_processor(
        self,
        prompt: Union[str, list[int]],
        mm_data_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
303
        tokenization_kwargs: Mapping[str, object],
304
305
306
        *,
        return_mm_hashes: bool,
    ) -> tuple[list[int], MultiModalKwargs, Optional[MultiModalHashes], bool]:
307
308
309
310
311
312
        (
            prompt_ids,
            mm_kwargs,
            mm_hashes,
            _,
        ) = super()._cached_apply_hf_processor(
313
314
315
            prompt=prompt,
            mm_data_items=mm_data_items,
            hf_processor_mm_kwargs=hf_processor_mm_kwargs,
316
            tokenization_kwargs=tokenization_kwargs,
317
            return_mm_hashes=return_mm_hashes,
318
319
320
        )

        # NOTE: The tokens are already inserted by the chat template
321
        return prompt_ids, mm_kwargs, mm_hashes, True
Patrick von Platen's avatar
Patrick von Platen committed
322

323
324
325
326

@MULTIMODAL_REGISTRY.register_processor(PixtralMultiModalProcessor,
                                        info=PixtralProcessingInfo,
                                        dummy_inputs=PixtralDummyInputsBuilder)
327
328
class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal,
                                      SupportsPP):
Patrick von Platen's avatar
Patrick von Platen committed
329

330
331
332
333
334
335
336
    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return None

        raise ValueError("Only image modality is supported")

337
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
Patrick von Platen's avatar
Patrick von Platen committed
338
        super().__init__()
339
340
        config = vllm_config.model_config.hf_config
        multimodal_config = vllm_config.model_config.multimodal_config
Patrick von Platen's avatar
Patrick von Platen committed
341
342
343
344
345
346
347
348
349
350
351
352
353
354
        self.config = config
        self.multimodal_config = multimodal_config

        dataclass_fields = {field.name for field in fields(VisionEncoderArgs)}
        vision_args = {
            key: value
            for key, value in self.config.vision_config.to_dict().items()
            if key in dataclass_fields
        }

        self.vision_args = VisionEncoderArgs(**vision_args)

        # init MistralForCausalLM
        self.language_model = init_vllm_registered_model(
355
            vllm_config=vllm_config,
356
357
358
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
        )
Patrick von Platen's avatar
Patrick von Platen committed
359
360

        self.vision_encoder = VisionTransformer(self.vision_args)
Patrick von Platen's avatar
Patrick von Platen committed
361
362
363
364
365
366
367
368
369
370
371
372

        if self.vision_args.add_pre_mm_projector_layer_norm:
            self.pre_mm_projector_norm = RMSNorm(self.vision_args.hidden_size,
                                                 eps=1e-5)

        if self.vision_args.mm_projector_id == PATCH_MERGE:
            self.patch_merger = PatchMerger(
                vision_encoder_dim=self.vision_args.hidden_size,
                spatial_merge_size=self.vision_args.spatial_merge_size,
                use_mlp_bias=False,
            )

Patrick von Platen's avatar
Patrick von Platen committed
373
374
375
        self.vision_language_adapter = VisionLanguageAdapter(
            self.vision_args, dim=config.text_config.hidden_size)

376
377
378
        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[PixtralImagePixelInputs]:
        images = kwargs.pop("images", None)
        if images is None:
            return None

        if not isinstance(images, (torch.Tensor, list)):
            raise ValueError("Incorrect type of images. "
                             f"Got type: {type(images)}")

        return PixtralImagePixelInputs(
            type="pixel_values",
            images=flatten_bn(images),
        )

    def _process_image_input(
        self,
        image_input: PixtralImagePixelInputs,
    ) -> tuple[torch.Tensor, ...]:
        images = image_input["images"]
        image_features = self.vision_encoder(images)
        feature_sizes = [
            image_feature.shape[0] for image_feature in image_features
        ]
Patrick von Platen's avatar
Patrick von Platen committed
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
        image_features = torch.cat(image_features)
        if self.vision_args.add_pre_mm_projector_layer_norm:
            image_features = self.pre_mm_projector_norm(image_features)
        if self.vision_args.mm_projector_id == PATCH_MERGE:
            patch_size = self.vision_args.patch_size
            spatial_merge_size_square = self.vision_args.spatial_merge_size**2
            img_patch_dims = [(img.shape[1] // patch_size,
                               img.shape[2] // patch_size) for img in images]
            feature_sizes = [
                feature_size // spatial_merge_size_square
                for feature_size in feature_sizes
            ]
            image_features = self.patch_merger(image_features,
                                               image_sizes=img_patch_dims)
        image_embeds = self.vision_language_adapter(image_features)
418
419
420
        image_embeds = torch.split(image_embeds, feature_sizes)
        return image_embeds

421
422
423
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

424
425
    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
426
        image_input = self._parse_and_validate_image_input(**kwargs)
427
        if image_input is None:
428
            return []
429

430
        return self._process_image_input(image_input)
431
432
433
434

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
435
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
436
437
    ) -> torch.Tensor:
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
438
439
        if multimodal_embeddings is not None \
            and len(multimodal_embeddings) != 0:
440
            inputs_embeds = merge_multimodal_embeddings(
441
442
                input_ids,
                inputs_embeds,
443
                multimodal_embeddings,
444
445
                self.vision_args.image_token_id,
            )
446
447
        return inputs_embeds

Patrick von Platen's avatar
Patrick von Platen committed
448
449
450
451
452
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
453
        inputs_embeds: Optional[torch.Tensor] = None,
Patrick von Platen's avatar
Patrick von Platen committed
454
        **kwargs: object,
455
    ) -> Union[torch.Tensor, IntermediateTensors]:
456
        """Run forward pass for pixtral."""
457
458
        if intermediate_tensors is not None:
            inputs_embeds = None
Patrick von Platen's avatar
Patrick von Platen committed
459

460
461
462
463
464
465
466
        # 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)
            input_ids = None
Patrick von Platen's avatar
Patrick von Platen committed
467
468
469

        hidden_states = self.language_model.model(input_ids,
                                                  positions,
470
                                                  intermediate_tensors,
Patrick von Platen's avatar
Patrick von Platen committed
471
472
473
474
475
476
477
478
479
480
481
482
                                                  inputs_embeds=inputs_embeds)

        return hidden_states

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

483
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
Patrick von Platen's avatar
Patrick von Platen committed
484

485
        def is_vision_encoder_weights(weight: tuple[str, torch.Tensor]):
Patrick von Platen's avatar
Patrick von Platen committed
486
487
            return weight[0].startswith("vision_encoder")

488
        def is_vision_lang_adapter_weights(weight: tuple[str, torch.Tensor]):
Patrick von Platen's avatar
Patrick von Platen committed
489
490
            return weight[0].startswith("vision_language_adapter")

491
        def is_patch_merger(weight: tuple[str, torch.Tensor]):
Patrick von Platen's avatar
Patrick von Platen committed
492
493
            return weight[0].startswith("patch_merger")

494
        def is_pre_mm_projector_norm(weight: tuple[str, torch.Tensor]):
Patrick von Platen's avatar
Patrick von Platen committed
495
496
            return weight[0].startswith("pre_mm_projector_norm")

497
        # Get references to parameters for direct loading
Patrick von Platen's avatar
Patrick von Platen committed
498
        vision_encoder_dict = dict(self.vision_encoder.named_parameters())
Patrick von Platen's avatar
Patrick von Platen committed
499
500
501
502
503
        patch_merger_dict = dict(self.patch_merger.named_parameters(
        )) if self.vision_args.mm_projector_id == PATCH_MERGE else dict()
        pre_mm_projector_norm_dict = dict(
            self.pre_mm_projector_norm.named_parameters(
            )) if self.vision_args.add_pre_mm_projector_layer_norm else dict()
504
        vision_lang_adapter_dict = dict(
Patrick von Platen's avatar
Patrick von Platen committed
505
            self.vision_language_adapter.named_parameters())
506
507
508
509
510
511
512
513
514
515

        def llm_weights_generator():
            # Single pass over weights
            for name, w in weights:
                if is_vision_encoder_weights((name, w)):
                    # Load vision encoder weights directly
                    trimmed_name = '.'.join(name.split(".")[1:])
                    param = vision_encoder_dict[trimmed_name]
                    with torch.no_grad():
                        default_weight_loader(param, w)
Patrick von Platen's avatar
Patrick von Platen committed
516
517
518
519
520
521
522
523
524
525
526
527
                elif is_patch_merger((name, w)):
                    # Load vision patch merger weights directly
                    trimmed_name = '.'.join(name.split(".")[1:])
                    param = patch_merger_dict[trimmed_name]
                    with torch.no_grad():
                        default_weight_loader(param, w)
                elif is_pre_mm_projector_norm((name, w)):
                    # Load vision pre_mm_projector_norm weights directly
                    trimmed_name = '.'.join(name.split(".")[1:])
                    param = pre_mm_projector_norm_dict[trimmed_name]
                    with torch.no_grad():
                        default_weight_loader(param, w)
528
529
530
531
532
533
534
535
536
537
538
539
540
                elif is_vision_lang_adapter_weights((name, w)):
                    # Load vision-language adapter weights directly
                    trimmed_name = '.'.join(name.split(".")[1:])
                    param = vision_lang_adapter_dict[trimmed_name]
                    with torch.no_grad():
                        default_weight_loader(param, w)
                else:
                    # LLM weights: yield them to be loaded
                    # by language_model.load_weights
                    yield (name, w)

        # Now we call the language model load with the generator
        self.language_model.load_weights(llm_weights_generator())
Patrick von Platen's avatar
Patrick von Platen committed
541
542
543
544
545
546
547
548
549
550
551
552
553
554


# Vision encoder
@dataclass
class VisionEncoderArgs:
    hidden_size: int
    num_channels: int
    image_size: int
    patch_size: int
    intermediate_size: int
    num_hidden_layers: int
    num_attention_heads: int
    rope_theta: float  # for rope-2D
    image_token_id: int
555
    adapter_bias: bool = True
Patrick von Platen's avatar
Patrick von Platen committed
556
557
558
    spatial_merge_size: int = 1
    add_pre_mm_projector_layer_norm: bool = False
    mm_projector_id: str = ""
Patrick von Platen's avatar
Patrick von Platen committed
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610


def _reshape_for_broadcast(freqs_cis: torch.Tensor,
                           x: torch.Tensor) -> torch.Tensor:
    """
    freqs_cis: complex - (seq_len, head_dim / 2)
    x: complex - (bsz, seq_len, head_dim / 2)
    """
    ndim = x.ndim
    assert ndim > 1
    assert freqs_cis.shape == (x.shape[1], x.shape[-1]), (
        freqs_cis.shape,
        (x.shape[1], x.shape[-1]),
    )
    shape = [
        d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)
    ]
    return freqs_cis.view(*shape)


def precompute_freqs_cis_2d(
    dim: int,
    height: int,
    width: int,
    theta: float,
) -> torch.Tensor:
    """
    freqs_cis: 2D complex tensor of shape (height, width, dim // 2)
        to be indexed by (height, width) position tuples
    """
    # (dim / 2) frequency bases
    freqs = 1.0 / (theta**(torch.arange(0, dim, 2).float() / dim))

    h = torch.arange(height, device=freqs.device)
    w = torch.arange(width, device=freqs.device)

    freqs_h = torch.outer(h, freqs[::2]).float()
    freqs_w = torch.outer(w, freqs[1::2]).float()
    freqs_2d = torch.cat(
        [
            freqs_h[:, None, :].repeat(1, width, 1),
            freqs_w[None, :, :].repeat(height, 1, 1),
        ],
        dim=-1,
    )
    return torch.polar(torch.ones_like(freqs_2d), freqs_2d)


def apply_rotary_emb_vit(
    xq: torch.Tensor,
    xk: torch.Tensor,
    freqs_cis: torch.Tensor,
611
) -> tuple[torch.Tensor, torch.Tensor]:
Patrick von Platen's avatar
Patrick von Platen committed
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
    assert freqs_cis.dtype == torch.complex64
    freqs_cis = _reshape_for_broadcast(freqs_cis, xq_)
    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
    return xq_out.type_as(xq), xk_out.type_as(xk)


class FeedForward(nn.Module):

    def __init__(self, args: VisionEncoderArgs):
        super().__init__()
        assert args.intermediate_size is not None
        self.w1 = nn.Linear(args.hidden_size,
                            args.intermediate_size,
                            bias=False)
        self.w2 = nn.Linear(args.intermediate_size,
                            args.hidden_size,
                            bias=False)
        self.w3 = nn.Linear(args.hidden_size,
                            args.intermediate_size,
                            bias=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.w2(F.silu(self.w1(x)) * self.w3(x))


class Attention(nn.Module):

    def __init__(self, args: VisionEncoderArgs):
        super().__init__()
        self.args = args
        assert not args.hidden_size % args.num_attention_heads
        self.n_heads = args.num_attention_heads
        self.head_dim = args.hidden_size // args.num_attention_heads

        self.wq = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
        self.wk = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
        self.wv = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
        self.wo = nn.Linear(args.hidden_size, args.hidden_size, bias=False)

    def forward(
        self,
        x: torch.Tensor,
657
        mask: torch.Tensor,
Patrick von Platen's avatar
Patrick von Platen committed
658
659
660
661
662
663
664
665
666
667
        freqs_cis: torch.Tensor,
    ) -> torch.Tensor:
        batch, patches, _ = x.shape

        q, k, v = self.wq(x), self.wk(x), self.wv(x)
        q = q.reshape(batch, patches, self.n_heads, self.head_dim)
        k = k.reshape(batch, patches, self.n_heads, self.head_dim)
        v = v.reshape(batch, patches, self.n_heads, self.head_dim)

        q, k = apply_rotary_emb_vit(q, k, freqs_cis=freqs_cis)
668
        out = xops.memory_efficient_attention(q, k, v, attn_bias=mask)
Patrick von Platen's avatar
Patrick von Platen committed
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
        out = out.reshape(batch, patches, self.n_heads * self.head_dim)
        return self.wo(out)


class TransformerBlock(nn.Module):

    def __init__(self, args: VisionEncoderArgs):
        super().__init__()
        self.attention = Attention(args)
        self.feed_forward = FeedForward(args)
        self.attention_norm = RMSNorm(args.hidden_size, eps=1e-5)
        self.ffn_norm = RMSNorm(args.hidden_size, eps=1e-5)

    def forward(
        self,
        x: torch.Tensor,
685
        mask: torch.Tensor,
Patrick von Platen's avatar
Patrick von Platen committed
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
        freqs_cis: torch.Tensor,
    ) -> torch.Tensor:
        r = self.attention.forward(self.attention_norm(x),
                                   mask=mask,
                                   freqs_cis=freqs_cis)
        h = x + r
        r = self.feed_forward.forward(self.ffn_norm(h))
        out = h + r
        return out


class Transformer(nn.Module):

    def __init__(self, args: VisionEncoderArgs):
        super().__init__()
        self.layers = torch.nn.ModuleList()
        for _ in range(args.num_hidden_layers):
            self.layers.append(TransformerBlock(args))

    def forward(
        self,
        x: torch.Tensor,
708
        mask: torch.Tensor,
Patrick von Platen's avatar
Patrick von Platen committed
709
710
711
712
713
714
715
        freqs_cis: Optional[torch.Tensor],
    ) -> torch.Tensor:
        for layer in self.layers:
            x = layer(x, mask=mask, freqs_cis=freqs_cis)
        return x


716
def position_meshgrid(patch_embeds_list: list[torch.Tensor], ) -> torch.Tensor:
Patrick von Platen's avatar
Patrick von Platen committed
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
    positions = torch.cat([
        torch.stack(
            torch.meshgrid(
                torch.arange(p.shape[-2]),
                torch.arange(p.shape[-1]),
                indexing="ij",
            ),
            dim=-1,
        ).reshape(-1, 2) for p in patch_embeds_list
    ])
    return positions


class VisionTransformer(nn.Module):

    def __init__(self, args: VisionEncoderArgs):
        super().__init__()
        self.args = args
        self.patch_conv = nn.Conv2d(
            in_channels=args.num_channels,
            out_channels=args.hidden_size,
            kernel_size=args.patch_size,
            stride=args.patch_size,
            bias=False,
        )
        self.ln_pre = RMSNorm(args.hidden_size, eps=1e-5)
        self.transformer = Transformer(args)

        head_dim = self.args.hidden_size // self.args.num_attention_heads
        assert head_dim % 2 == 0, "ROPE requires even head_dim"
        self._freqs_cis: Optional[torch.Tensor] = None

    @property
    def max_patches_per_side(self) -> int:
        return self.args.image_size // self.args.patch_size

    @property
754
    def device(self) -> torch.types.Device:
Patrick von Platen's avatar
Patrick von Platen committed
755
756
757
        return next(self.parameters()).device

    @property
758
    def dtype(self) -> torch.dtype:
Patrick von Platen's avatar
Patrick von Platen committed
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
        return next(self.parameters()).dtype

    @property
    def freqs_cis(self) -> torch.Tensor:
        if self._freqs_cis is None:
            self._freqs_cis = precompute_freqs_cis_2d(
                dim=self.args.hidden_size // self.args.num_attention_heads,
                height=self.max_patches_per_side,
                width=self.max_patches_per_side,
                theta=self.args.rope_theta,
            )

        if self._freqs_cis.device != self.device:
            self._freqs_cis = self._freqs_cis.to(device=self.device)

        return self._freqs_cis

    def forward(
        self,
778
        images: list[torch.Tensor],
Patrick von Platen's avatar
Patrick von Platen committed
779
780
781
    ) -> torch.Tensor:
        """
        Args:
782
            images: list of N_img images of variable sizes,
Patrick von Platen's avatar
Patrick von Platen committed
783
784
                each of shape (C, H, W)
        Returns:
785
            image_features: tensor of token features for
Patrick von Platen's avatar
Patrick von Platen committed
786
787
788
789
790
791
792
                all tokens of all images of shape (N_toks, D)
        """
        # pass images through initial convolution independently
        patch_embeds_list = [
            self.patch_conv(img.unsqueeze(0).to(self.dtype)) for img in images
        ]

793
794
795
796
797
        patch_embeds = [
            p.flatten(2).permute(0, 2, 1) for p in patch_embeds_list
        ]
        embed_sizes = [p.shape[1] for p in patch_embeds]

Patrick von Platen's avatar
Patrick von Platen committed
798
        # flatten to a single sequence
799
        patch_embeds = torch.cat(patch_embeds, dim=1)
Patrick von Platen's avatar
Patrick von Platen committed
800
801
802
803
804
805
806
        patch_embeds = self.ln_pre(patch_embeds)

        # positional embeddings
        positions = position_meshgrid(patch_embeds_list).to(self.device)
        freqs_cis = self.freqs_cis[positions[:, 0], positions[:, 1]]

        # pass through Transformer with a block diagonal mask delimiting images
807
808
809
810
811
812
        if USE_XFORMERS_OPS:
            mask = xops.fmha.attn_bias.BlockDiagonalMask.from_seqlens(
                [p.shape[-2] * p.shape[-1] for p in patch_embeds_list], )
        else:
            raise ImportError("Xformers is required for Pixtral inference "
                              "with the Mistral format")
Patrick von Platen's avatar
Patrick von Platen committed
813
814
        out = self.transformer(patch_embeds, mask=mask, freqs_cis=freqs_cis)

815
816
        # squeeze dim 0 and split into separate tensors for each image
        return torch.split(out.squeeze(0), embed_sizes)
Patrick von Platen's avatar
Patrick von Platen committed
817
818
819
820
821
822
823
824
825
826


class VisionLanguageAdapter(nn.Module):

    def __init__(self, args: VisionEncoderArgs, dim: int):
        super().__init__()
        assert isinstance(args, VisionEncoderArgs)
        self.w_in = nn.Linear(
            args.hidden_size,
            dim,
827
            bias=args.adapter_bias,
Patrick von Platen's avatar
Patrick von Platen committed
828
829
        )
        self.gelu = nn.GELU()
830
        self.w_out = nn.Linear(dim, dim, bias=args.adapter_bias)
Patrick von Platen's avatar
Patrick von Platen committed
831
832
833

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.w_out(self.gelu(self.w_in(x)))
834
835


Patrick von Platen's avatar
Patrick von Platen committed
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
class PatchMerger(nn.Module):
    """
    Learned merging of spatial_merge_size ** 2 patches
    """

    def __init__(
        self,
        vision_encoder_dim: int,
        spatial_merge_size: int,
        use_mlp_bias: bool = False,
    ) -> None:
        super().__init__()

        mlp_input_dim = vision_encoder_dim * (spatial_merge_size**2)

        self.spatial_merge_size = spatial_merge_size
        self.mlp_input_dim = mlp_input_dim

        self.merging_layer = nn.Linear(
            mlp_input_dim,
            vision_encoder_dim,
            bias=use_mlp_bias,
        )

    def forward(self, x: torch.Tensor,
                image_sizes: list[tuple[int, int]]) -> torch.Tensor:
        # image_sizes specified in tokens
        assert sum([h * w for h, w in image_sizes]) == len(x)

        # x is (N, vision_encoder_dim)
        x = self.permute(x, image_sizes)

868
869
        # x is (N / spatial_merge_size ** 2,
        #       vision_encoder_dim * spatial_merge_size ** 2)
Patrick von Platen's avatar
Patrick von Platen committed
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
        x = self.merging_layer(x)

        # x is (N / spatial_merge_size ** 2, vision_encoder_dim)
        return x

    def permute(
        self,
        x: torch.Tensor,
        image_sizes: list[tuple[int, int]],
    ) -> torch.Tensor:
        """
        Args:
            x: (N, D) where N is flattened and concatenated patch tokens
                for all images
            image_sizes: list of tuple of (height, width) in tokens for
                each image
        Returns:
            image_features: reorders patch tokens so each grid of
                (spatial_merge_size, spatial_merge_size) is contiguous.
                now (N / spatial_merge_size ** 2, D * spatial_merge_size ** 2)
        """

        sub_grids = get_sub_grids(
            x=x,
            image_sizes=image_sizes,
            spatial_merge_size=self.spatial_merge_size
        )  # list of [d x sub_grid_size x sub_grid_size x n_patches]
        permuted_tensor: list[torch.Tensor] = []
        for grid in sub_grids:
            n_patches = grid.shape[-1]
            permuted_tensor.append(grid.view(-1, n_patches).t(
            ))  # n_patches x d * sub_grid_size * sub_grid_size
        return torch.cat(
            permuted_tensor, dim=0
        )  # (N / spatial_merge_size ** 2, d * spatial_merge_size ** 2)


def get_sub_grids(
    x: torch.Tensor,
    image_sizes: list[tuple[int, int]],
    spatial_merge_size: int,
) -> list[torch.Tensor]:
    # image_sizes specified in tokens
    tokens_per_image = [h * w for h, w in image_sizes]
    d = x.shape[-1]
    all_img_sub_grids: list[torch.Tensor] = []
    sub_grid_size = spatial_merge_size

    for image_index, image_tokens in enumerate(x.split(tokens_per_image)):
        # Reshape image_tokens into a 2D grid
        h, w = image_sizes[image_index]
        image_grid = image_tokens.view(h, w, d).permute(
            2, 0, 1)[None, :, :, :]  # 1 x d x h x w
        sub_grids = torch.nn.functional.unfold(image_grid,
                                               kernel_size=sub_grid_size,
                                               stride=sub_grid_size)
        sub_grids = sub_grids.view(
            1, d, sub_grid_size, sub_grid_size,
            -1)  # 1 x d x sub_grid_size x sub_grid_size x n_patches

        all_img_sub_grids.append(sub_grids[0])

    return all_img_sub_grids


935
936
937
938
939
940
941
942
#### HF Transformers version of Pixtral ####
# Based off https://github.com/huggingface/transformers/blob/d7950bff82b18c823193d17d72188c5e46d06c83/src/transformers/models/pixtral/modeling_pixtral.py
# This model follows the Llava family, meaning image embeddings are placed
# instead of the `[IMG]` token placeholders.
# The model uses [`PixtralVisionModel`] for its vision encoder,
# and [`MistralForCausalLM`] for its language decoder.


943
944
945
946
947
948
949
950
class PixtralHFEncoderInfo(VisionEncoderInfo[PixtralVisionConfig]):

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> int:
951
952
953
        ncols, nrows = self.get_patch_grid_size(
            image_width=image_width,
            image_height=image_height,
954
        )
955
        return ncols * nrows
956

957
958
959
960
    def get_image_size(self) -> int:
        return self.vision_config.image_size

    def get_patch_size(self) -> int:
961
962
963
        # spatial_merge_size is needed for Mistral3
        spatial_merge_size = getattr(self.hf_config, "spatial_merge_size", 1)
        return self.vision_config.patch_size * spatial_merge_size
964
965

    def get_patch_grid_length(self) -> int:
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
        image_size, patch_size = self.get_image_size(), self.get_patch_size()

        # Since interpolation is applied, the image size need not be divisible
        # assert image_size % patch_size == 0
        return image_size // patch_size

    # Adapted from: https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/pixtral/image_processing_pixtral.py#L99
    def get_patch_grid_size(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> tuple[int, int]:
        max_width = max_height = self.get_image_size()
        patch_width = patch_height = self.get_patch_size()

        ratio = max(image_width / max_width, image_height / max_height)

        if ratio > 1:
985
986
            image_width = int(math.floor(image_width / ratio))
            image_height = int(math.floor(image_height / ratio))
987
988
989
990
991
992
993

        nrows, ncols = _get_pixtral_hf_num_image_tokens(
            (image_height, image_width),
            (patch_height, patch_width),
        )  # type: ignore

        return ncols, nrows
994
995
996
997


class PixtralHFMLP(nn.Module):

998
999
1000
1001
1002
1003
1004
    def __init__(
        self,
        config: PixtralVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        prefix: str = "",
    ) -> None:
1005
        super().__init__()
1006

1007
        assert config.intermediate_size is not None
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=config.hidden_size,
            output_sizes=[config.intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
        self.down_proj = RowParallelLinear(input_size=config.intermediate_size,
                                           output_size=config.hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.down_proj")
        self.act_and_mul = get_act_and_mul_fn(config.hidden_act)
1020
1021

    def forward(self, x: torch.Tensor) -> torch.Tensor:
1022
1023
1024
1025
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_and_mul(gate_up)
        x, _ = self.down_proj(x)
        return x
1026
1027
1028
1029


class PixtralHFAttention(nn.Module):

1030
1031
1032
1033
1034
1035
1036
    def __init__(
        self,
        config: PixtralVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        prefix: str = "",
    ) -> None:
1037
        super().__init__()
1038

1039
1040
        self.config = config
        assert not config.hidden_size % config.num_attention_heads
1041
1042
1043
        self.total_num_heads = config.num_attention_heads
        tp_size = get_tensor_model_parallel_world_size()
        self.n_heads = divide(config.num_attention_heads, tp_size)
1044
1045
        self.head_dim = config.hidden_size // config.num_attention_heads

1046
1047
1048
        self.qkv_proj = QKVParallelLinear(
            hidden_size=config.hidden_size,
            head_size=self.head_dim,
1049
            total_num_heads=self.total_num_heads,
1050
1051
1052
1053
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
1054
        assert self.total_num_heads * self.head_dim == config.hidden_size
1055
1056
1057
1058
1059
1060
1061
        self.o_proj = RowParallelLinear(
            input_size=config.hidden_size,
            output_size=config.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )
1062
1063
1064
1065

    def forward(
        self,
        hidden_states: torch.Tensor,
1066
        attention_mask: torch.Tensor,
1067
        position_embeddings: torch.Tensor,
1068
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
1069
        batch, patches, _ = hidden_states.size()
1070

1071
1072
        qkv_states, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv_states.chunk(3, dim=-1)
1073

1074
1075
1076
        # Transpose q and k to apply HF's Rotary Position Embedding
        q = q.view(batch, patches, self.n_heads, self.head_dim).transpose(1, 2)
        k = k.view(batch, patches, self.n_heads, self.head_dim).transpose(1, 2)
1077
        v = v.view(batch, patches, self.n_heads, self.head_dim)
1078
        cos, sin = position_embeddings
1079
        q, k = apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=0)
1080

1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
        if USE_XFORMERS_OPS:
            # Transpose q and k back for attention
            q = q.transpose(1, 2).contiguous()
            k = k.transpose(1, 2).contiguous()

            out = xops.memory_efficient_attention(q,
                                                  k,
                                                  v,
                                                  attn_bias=attention_mask)
        else:
1091
            v = v.transpose(1, 2)
1092
1093
1094
            out = nn.functional.scaled_dot_product_attention(
                q, k, v, attn_mask=attention_mask)
            out = out.transpose(1, 2)
1095

1096
1097
        out = out.view(batch, patches, self.n_heads * self.head_dim)
        attn_output, _ = self.o_proj(out)
1098

1099
        return attn_output, None
1100
1101
1102
1103


class PixtralHFTransformerBlock(nn.Module):

1104
1105
1106
1107
1108
1109
1110
    def __init__(
        self,
        config: PixtralVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        prefix: str = "",
    ) -> None:
1111
        super().__init__()
1112

1113
        self.attention_norm = RMSNorm(config.hidden_size, eps=1e-5)
1114
1115
1116
1117
1118
1119
        self.attention = PixtralHFAttention(config,
                                            quant_config=quant_config,
                                            prefix=f"{prefix}.attention")
        self.feed_forward = PixtralHFMLP(config,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.feed_forward")
1120
1121
1122
1123
1124
        self.ffn_norm = RMSNorm(config.hidden_size, eps=1e-5)

    def forward(
        self,
        hidden_states: torch.Tensor,
1125
        attention_mask: torch.Tensor,
1126
1127
        position_embeddings: torch.Tensor,
    ) -> torch.Tensor:
1128
1129
1130
        r, _ = self.attention.forward(self.attention_norm(hidden_states),
                                      attention_mask=attention_mask,
                                      position_embeddings=position_embeddings)
1131
1132
1133
1134
1135
1136
1137
1138
        h = hidden_states + r
        r = self.feed_forward.forward(self.ffn_norm(h))
        out = h + r
        return out


class PixtralHFTransformer(nn.Module):

1139
1140
1141
1142
1143
1144
1145
1146
    def __init__(
        self,
        config: PixtralVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        prefix: str = "",
    ) -> None:
1147
        super().__init__()
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159

        if num_hidden_layers_override is None:
            num_hidden_layers = config.num_hidden_layers
        else:
            num_hidden_layers = num_hidden_layers_override

        self.layers = nn.ModuleList([
            PixtralHFTransformerBlock(config=config,
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.layers.{layer_idx}")
            for layer_idx in range(num_hidden_layers)
        ])
1160
1161
1162
1163

    def forward(
        self,
        x: torch.Tensor,
1164
        attention_mask: torch.Tensor,
1165
        position_embeddings: torch.Tensor,
1166
        return_all_hidden_states: bool,
1167
    ) -> torch.Tensor:
1168
        hidden_states_pool = [x]
1169

1170
1171
        for layer in self.layers:
            x = layer(x, attention_mask, position_embeddings)
1172
1173
1174
1175
1176
1177
            if return_all_hidden_states:
                hidden_states_pool.append(x)
        # If we have multiple feature sample layers, we return all hidden
        # states in order and grab the ones we need by index.
        if return_all_hidden_states:
            return hidden_states_pool
1178
1179
1180
1181
1182
        return x


class PixtralHFVisionModel(nn.Module):

1183
1184
1185
1186
1187
1188
1189
1190
1191
    def __init__(
        self,
        config: PixtralVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        require_post_norm: Optional[bool] = None,
        prefix: str = "",
    ) -> None:
1192
1193
1194
        super().__init__()

        self.config = config
1195

1196
1197
1198
1199
1200
1201
1202
1203
        self.patch_conv = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=config.hidden_size,
            kernel_size=config.patch_size,
            stride=config.patch_size,
            bias=False,
        )
        self.ln_pre = RMSNorm(config.hidden_size, eps=1e-5)
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
        self.transformer = PixtralHFTransformer(
            config,
            quant_config,
            num_hidden_layers_override=num_hidden_layers_override,
            prefix=f"{prefix}.transformer",
        )

        num_hidden_layers = config.num_hidden_layers
        if len(self.transformer.layers) > config.num_hidden_layers:
            raise ValueError(
                f"The original encoder only has {num_hidden_layers} "
                f"layers, but you requested {len(self.transformer.layers)} "
                "layers.")

        if require_post_norm is True:
            msg = "PixtralHFVisionModel does not have post-layernorm"
            raise ValueError(msg)

1222
1223
1224
1225
1226
1227
1228
        self.dtype = next(self.parameters()).dtype
        self.device = next(self.parameters()).device
        self.patch_positional_embedding = PixtralRotaryEmbedding(
            config, self.device)

    def forward(
        self,
1229
        pixel_values: list[torch.Tensor],
1230
        feature_sample_layers: Optional[list[int]] = None,
1231
    ) -> tuple[torch.Tensor, ...]:
1232
1233
        """
        Args:
1234
1235
1236
1237
            pixel_values: Each image to be processed will be a separate tensor
                in pixel_values. This means it will be a list of tensors
                because multiple requests batched can have multiple images,
                each with their own shape potentially
1238
1239
1240
            feature_sample_layers: Layer indices whose features should be
                concatenated and used as the visual encoder output. If none
                are provided, the last layer is used.
1241

1242
1243
1244
1245
1246
1247
        Returns:
            image_features: tensor of token features for
                all tokens of all images of shape (N_toks, D)
        """
        # pass images through initial convolution independently
        patch_embeds_list = [
1248
            self.patch_conv(img.unsqueeze(0).to(self.dtype))
1249
1250
1251
            for img in pixel_values
        ]

1252
1253
1254
1255
1256
        patch_embeds = [
            p.flatten(2).permute(0, 2, 1) for p in patch_embeds_list
        ]
        embed_sizes = [p.shape[1] for p in patch_embeds]

1257
        # flatten to a single sequence
1258
        patch_embeds = torch.cat(patch_embeds, dim=1)
1259
1260
1261
1262
1263
1264
1265
1266
1267
        patch_embeds = self.ln_pre(patch_embeds)

        # positional embeddings
        position_ids = position_ids_in_meshgrid(
            patch_embeds_list,
            max_width=self.config.image_size // self.config.patch_size).to(
                self.device)
        position_embedding = self.patch_positional_embedding(
            patch_embeds, position_ids)
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278

        if USE_XFORMERS_OPS:
            attention_mask = xops.fmha.attn_bias.BlockDiagonalMask.from_seqlens(
                [p.shape[-2] * p.shape[-1] for p in patch_embeds_list], )
        else:
            from transformers.models.pixtral.modeling_pixtral import (
                generate_block_attention_mask)
            attention_mask = generate_block_attention_mask(
                [p.shape[-2] * p.shape[-1] for p in patch_embeds_list],
                patch_embeds)

1279
1280
1281
1282
1283
1284
1285
1286
1287
        return_all_hidden_states = feature_sample_layers is not None
        out = self.transformer(
            patch_embeds,
            attention_mask,
            position_embedding,
            return_all_hidden_states=return_all_hidden_states)

        out = resolve_visual_encoder_outputs(out, feature_sample_layers, None,
                                             self.config.num_hidden_layers)
1288

1289
        # squeeze dim 0 and split into separate tensors for each image
1290
        return torch.split(out.squeeze(0), embed_sizes)
1291
1292
1293

    # (TODO) Add prefix argument for filtering out weights to be loaded
    #        ref: https://github.com/vllm-project/vllm/pull/7186#discussion_r1734163986
1294
1295
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
1296
1297
1298
1299
1300
1301
1302
1303
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
1304
        params_dict = dict(self.named_parameters())
1305
        loaded_params: set[str] = set()
1306
        layer_count = len(self.transformer.layers)
1307
1308

        for name, loaded_weight in weights:
1309
1310
1311
1312
1313
1314
            # omit layers when num_hidden_layers_override is set
            if name.startswith("transformer.layers"):
                layer_idx = int(name.split(".")[2])
                if layer_idx >= layer_count:
                    continue

1315
1316
1317
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
1318
1319
                name = name.replace(weight_name, param_name)
                param = params_dict[name]
1320
1321
1322
1323
1324
1325
1326
1327
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
1328
1329
            loaded_params.add(name)
        return loaded_params