File size: 35,271 Bytes
e9f0a60
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
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
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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
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
404
405
406
407
408
409
410
411
412
413
414
415
416
417
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
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
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
611
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
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
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
from typing import List, Optional, Tuple, Union
from dataclasses import dataclass

import torch
from torch import nn
from transformers import (
    AutoConfig,
    AutoModel,
    AutoModelForCausalLM,
    LlavaNextForConditionalGeneration,
    LlavaNextModel,
)
from transformers.models.llava_next.modeling_llava_next import (
    LlavaNextCausalLMOutputWithPast,
    LlavaNextPreTrainedModel,
    LlavaNextMultiModalProjector,
    get_anyres_image_grid_shape,
    image_size_to_num_patches,
    unpad_image,
    LlavaNextModelOutputWithPast
)
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, can_return_tuple, logging
from accelerate import init_empty_weights
from transformers import Blip2QFormerConfig, Blip2QFormerModel
from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig
from .configuration import Granite4VisionConfig, Granite4VisionConfigNaflex
from .downsampling import BilinearDownsampler, QFormerDownsampler, WindowQFormerDownsampler
import math
import numpy as np
from fractions import Fraction
from transformers.modeling_utils import flash_attention_forward
from transformers.models.granitemoehybrid.modeling_granitemoehybrid import HybridMambaAttentionDynamicCache
IGNORE_INDEX = -100
logger = logging.get_logger(__name__)

@dataclass
class Granite4VisionModelOutputWithPast(LlavaNextModelOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    balancing_loss: Optional[torch.FloatTensor] = None

@dataclass
class Granite4VisionCausalLMOutputWithPast(LlavaNextCausalLMOutputWithPast):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    image_hidden_states (`torch.FloatTensor`, *optional*):
        A `torch.FloatTensor` of size `(batch_size, num_images, sequence_length, hidden_size)`.
        image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
    """

    balancing_loss: Optional[torch.FloatTensor] = None


class ParamWrapper(nn.Module):
    def __init__(self, param):
        super().__init__()
        self.param = param

class Granite4VisionForConditionalGeneration(LlavaNextForConditionalGeneration):
    config_class = Granite4VisionConfig

    def __init__(self, config: Granite4VisionConfig):
        # Update config with pretrained models if specified
        if config.pretrained_vision_tower:
            config.vision_config = AutoConfig.from_pretrained(
                config.pretrained_vision_tower, **config.vision_config.to_dict()
            )
            config.vision_config = (
                config.vision_config.vision_config
                if hasattr(config.vision_config, "vision_config")
                else config.vision_config
            )
        if config.pretrained_language_model:
            config.text_config = AutoConfig.from_pretrained(
                config.pretrained_language_model, **config.text_config.to_dict()
            )

        # Initialize parent
        LlavaNextPreTrainedModel.__init__(self, config)

        # Create custom model instance
        self.model = Granite4VisionModel(config)

        # Create lm_head
        self.lm_head = nn.Linear(
            config.text_config.hidden_size, config.text_config.vocab_size, bias=False
        )

        # Load pretrained components if specified
        if config.pretrained_vision_tower:
            self._load_pretrained_vision_tower(config)
            config.pretrained_vision_tower = ""

        if config.pretrained_language_model:
            self._load_pretrained_language_model(config)
            config.pretrained_language_model = ""

        self.post_init()

    def _load_pretrained_vision_tower(self, config):
        """Load pretrained vision tower weights"""
        print(f"Loading vision tower from: {config.pretrained_vision_tower}")
        vision_tower = AutoModel.from_pretrained(
            config.pretrained_vision_tower,
            attn_implementation="flash_attention_2",
            device_map="cpu",
            dtype=torch.bfloat16,

        )
        self.model.vision_tower = self.model.vision_tower.to(torch.bfloat16)
        print(self.model.vision_tower.load_state_dict(vision_tower.state_dict(), strict=False).missing_keys)
        self.model.vision_tower.config._attn_implementation = "flash_attention_2"
        # todo: (Avihu) would have done this but afraid - maybe something I'm missing
        # self.model.vision_tower = vision_tower
        self.config.vision_config = (
            self.model.vision_tower.config.vision_config
            if hasattr(self.model.vision_tower.config, "vision_config")
            else self.model.vision_tower.config
        )

    def _load_pretrained_language_model(self, config):
        """Load pretrained language model weights"""
        print(f"Loading language model from: {config.pretrained_language_model}")
        language_model = AutoModelForCausalLM.from_pretrained(
            config.pretrained_language_model,
            device_map="cpu",
            attn_implementation="flash_attention_2",
            dtype=torch.bfloat16,
            # use_kernels=True,
        )
        if self.config.image_token_index >= language_model.config.vocab_size:
            language_model.resize_token_embeddings(self.config.image_token_index + 1)
        # load weights in quantized mode with kernels
        self.model.language_model = language_model.model
        self.lm_head = language_model.lm_head
        # Load weights into the language model inside self.model
        self.config.text_config = self.model.language_model.config


    @can_return_tuple
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_sizes: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        vision_feature_layer: Optional[Union[int, list[int]]] = None,
        vision_feature_select_strategy: Optional[str] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        spatial_shapes: Optional[torch.LongTensor] = None,
        pixel_attention_mask: Optional[torch.Tensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> Union[tuple, Granite4VisionCausalLMOutputWithPast]:

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )

        outputs = self.model(
            input_ids,
            pixel_values=pixel_values,
            image_sizes=image_sizes,
            vision_feature_layer=vision_feature_layer,
            vision_feature_select_strategy=vision_feature_select_strategy,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            cache_position=cache_position,
            spatial_shapes=spatial_shapes,
            pixel_attention_mask=pixel_attention_mask,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        logits = logits / self.config.text_config.logits_scaling

        loss = None
        if labels is not None:
            # Upcast to float if we need to compute the loss to avoid potential precision issues
            # Avihu: removed the .float(), didn't make a huge difference and requires more memory
            # logits = logits.float()
            # Flatten the tokens
            loss = self.loss_function(
                logits,
                labels,
                vocab_size=self.config.text_config.vocab_size,
                **kwargs,
            )

        return Granite4VisionCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=outputs.image_hidden_states,
            balancing_loss=outputs.balancing_loss
        )
        
    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        inputs_embeds=None,
        pixel_values=None,
        image_sizes=None,
        attention_mask=None,
        cache_position=None,
        logits_to_keep=None,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            logits_to_keep=logits_to_keep,
            **kwargs,
        )

        # Check if the model or its langauge model are moe type - requires special cache handling
        if any(class_name in self.__class__.__name__.lower() or class_name in self.language_model.__class__.__name__.lower() for class_name in ["moe"]):
            model_inputs = self.prepare_inputs_for_generation_granite_moe(**model_inputs)
            
        # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
        # Otherwise we need pixel values to be passed to model
        if cache_position[0] == 0:
            model_inputs["pixel_values"] = pixel_values
            model_inputs["image_sizes"] = image_sizes

        return model_inputs

    # Avihu: would have used the GraniteMoeSharedForCausalLM method, but we don't store this object anymore (split the model / lm head)
    def prepare_inputs_for_generation_granite_moe(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        **kwargs,
    ):
        # Overwritten -- has a unique cache type, `HybridMambaAttentionDynamicCache`
        # Note: (Avihu) in transformers v4, the past_key_values is already an empty DynamicCache object. Testing that too
        empty_past_kv = past_key_values is None or (isinstance(past_key_values, DynamicCache) and past_key_values[0][0] is None)

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
        #              (we can't check exception 3 while compiling)
        if not empty_past_kv:
            if (
                inputs_embeds is not None  # Exception 1
                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
            ):
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]
        elif use_cache:
            past_key_values = HybridMambaAttentionDynamicCache(
                self.model.language_model.config, input_ids.shape[0], self.dtype, device=self.device
            )

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if not empty_past_kv:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and empty_past_kv:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
            }
        )

        # Forward ALL kwargs that are uninitialized (e.g. `use_cache`).
        for key, value in kwargs.items():
            if key not in model_inputs:
                model_inputs[key] = value

        return model_inputs


class Granite4VisionModel(LlavaNextPreTrainedModel):
    config_class = Granite4VisionConfig
    
    def __init__(self, config: Granite4VisionConfig):
        super().__init__(config)
        self.vision_tower = AutoModel.from_config(config.vision_config)
        self.multi_modal_projector = LlavaNextMultiModalProjector(config)
        self.downsampler = None
        self.downsample_rate = config.downsample_rate
        if config.downsample_rate is not None:
            if config.downsample_method in ["interpolate", "bilinear"]:
                self.downsampler = BilinearDownsampler(config)
            elif config.downsample_method == "qformer":
                self.downsampler = QFormerDownsampler(config)
            elif config.downsample_method == "window_qformer":
                self.downsampler = WindowQFormerDownsampler(config)
        self.image_newline = None
        if config.use_image_newline_parameter:
            embed_std = 1 / math.sqrt(config.text_config.hidden_size)
            image_newline = nn.Parameter(torch.randn(config.text_config.hidden_size, dtype=self.dtype) * embed_std)
            self.model_type = config.model_type
            if self.model_type in ["gpt_vision", "granite4_vision"]:
                # this hack allows to do lora training from scratch, so image_newline would be in modules_to_keep
                self.image_newline = ParamWrapper(image_newline)
            else:
                self.image_newline = image_newline
        self.vocab_size = config.text_config.vocab_size
        
        # with init_empty_weights(): # Avihu: hack to load the model faster
        self.language_model = AutoModel.from_config(config.text_config)
        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
        self.post_init()
        
    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def set_decoder(self, decoder):
        self.language_model = decoder

    def get_decoder(self):
        return self.language_model


    def pack_image_features(self, image_features, image_sizes, vision_feature_select_strategy, image_newline=None):
        """
        Reshape, unpad and then pack each image_feature into a single image_features tensor containing all visual vectors.

        Args:
            image_features (`list[torch.Tensor]` of length num_images, each of shape `(num_patches, image_length, embed_dim)`)
                List of image feature tensor, each contains all the visual feature of all patches.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_select_strategy (`str`)
                The feature selection strategy used to select the vision feature from the vision backbone.
            image_newline (`torch.Tensor` of shape `(embed_dim)`)
                New line embedding vector.
        Returns:
            image_features (`torch.Tensor` of shape `(all_feat_len, embed_dim)`)
            feature_lens (`list[int]`)
                token length of each image in image_features
        """
        new_image_features = []
        feature_lens = []
        for image_idx, image_feature in enumerate(image_features):
            if image_feature.shape[0] > 1:
                base_image_feature = image_feature[0]
                image_feature = image_feature[1:]
                height = width = self.config.vision_config.image_size // self.config.vision_config.patch_size

                num_patch_height, num_patch_width = get_anyres_image_grid_shape(
                    image_sizes[image_idx],
                    self.config.image_grid_pinpoints,
                    self.config.vision_config.image_size,
                )
                if self.downsampler is not None:
                    ds_rate = Fraction(self.downsample_rate)
                    height = int(height  * ds_rate)
                    width = int(width  * ds_rate)

                if (
                    np.prod(image_feature.shape) % (num_patch_height * num_patch_width * height * width) != 0
                    and vision_feature_select_strategy == "default"
                ):
                    logger.warning_once(
                        "Image feature shape does not line up with the provided patch size. "
                        "You may be using the `default` vision_feature_select_strategy with a"
                        " visual encoder that does not have CLS."
                    )

                image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
                image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
                image_feature = image_feature.flatten(1, 2).flatten(2, 3)
                image_feature = unpad_image(image_feature, image_sizes[image_idx])
                if image_newline is not None:
                    image_feature = torch.cat(
                        (
                            image_feature,
                            image_newline[:, None, None]
                            .expand(*image_feature.shape[:-1], 1)
                            .to(image_feature.device, image_feature.dtype),
                        ),
                        dim=-1,
                    )
                image_feature = image_feature.flatten(1, 2).transpose(0, 1)
                image_feature = torch.cat((base_image_feature, image_feature), dim=0)
            else:
                image_feature = image_feature[0]
                if image_newline is not None:
                    image_feature = torch.cat((image_feature, image_newline[None].to(image_feature)), dim=0)
            new_image_features.append(image_feature)
            feature_lens.append(image_feature.size(0))
        feature_lens = torch.tensor(feature_lens, dtype=torch.long, device=image_features[0].device)
        return new_image_features, feature_lens

    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_sizes: torch.Tensor,
        vision_feature_layer: Optional[Union[int, list[int]]] = None,
        vision_feature_select_strategy: Optional[str] = None,
    ):
        """
        Obtains image last hidden states from the vision tower and apply multimodal projection.

        Args:
            pixel_values (`torch.FloatTensor]` of shape `(batch_size, num_patches, channels, height, width)`)
               The tensors corresponding to the input images.
            image_sizes (`torch.Tensor` of shape `(num_images, 2)`)
                Actual image size of each images (H, W).
            vision_feature_layer (`Union[int, list[int]]`, *optional*):
                The index of the layer to select the vision feature. If multiple indices are provided,
                the vision feature of the corresponding indices will be concatenated to form the
                vision features.
            vision_feature_select_strategy (`str`, *optional*):
                The feature selection strategy used to select the vision feature from the vision backbone.
                Can be one of `"default"` or `"full"`
        Returns:
            image_features (list[`torch.Tensor`]): List of image feature tensor, each contains all the visual feature of all patches
            and are of shape `(num_patches, image_length, embed_dim)`).
        """
        vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )

        # ! infer image_num_patches from image_sizes
        image_num_patches = [
            image_size_to_num_patches(
                image_size=imsize,
                grid_pinpoints=self.config.image_grid_pinpoints,
                patch_size=self.config.vision_config.image_size,
            )
            for imsize in image_sizes
        ]
        if pixel_values.dim() == 5:
            # stacked if input is (batch_size, num_patches, num_channels, height, width)
            _pixel_values_list = [pix_val[:num_patch] for pix_val, num_patch in zip(pixel_values, image_num_patches)]
            pixel_values = torch.cat(_pixel_values_list, dim=0)
        elif pixel_values.dim() != 4:
            # otherwise has to be stacked from list of (num_patches, num_channels, height, width)
            raise ValueError(f"pixel_values of shape {pixel_values.shape}, expect to be of 4 or 5 dimensions")

        image_features = self.vision_tower(pixel_values, output_hidden_states=True)
        # If we have one vision feature layer, return the corresponding hidden states,
        # otherwise, select the hidden states of each feature layer and concatenate them
        if isinstance(vision_feature_layer, int):
            selected_image_feature = image_features.hidden_states[vision_feature_layer]
        else:
            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
            selected_image_feature = torch.cat(hs_pool, dim=-1)

        if vision_feature_select_strategy == "default":
            selected_image_feature = selected_image_feature[:, 1:]

        image_features = self.multi_modal_projector(selected_image_feature)
        if self.downsampler is not None:
            # training this with peft+deepspeed had this issue+fix: 
            # https://github.com/deepspeedai/DeepSpeed/issues/7203#issuecomment-3007490737
            image_features = self.downsampler(image_features)
        if image_features.shape[0] != sum(image_num_patches):
            print("about to crash on split", pixel_values.shape, image_sizes, image_num_patches)
        image_features = torch.split(image_features, image_num_patches, dim=0)

        # NOTE we only support multimodal_patch_merge_type == "spatial_unpad"
        image_newline = self.image_newline.param if self.model_type in ["gpt_vision", "granite4_vision"] else self.image_newline
        image_features, feature_lens = self.pack_image_features(
            image_features,
            image_sizes,
            vision_feature_select_strategy=vision_feature_select_strategy,
            image_newline=image_newline,
        )
        return image_features

    def get_image_features_naflex(
        self,
        pixel_values: torch.FloatTensor,
        spatial_shapes, 
        pixel_attention_mask,
        vision_feature_layer: Optional[Union[int, list[int]]] = None,
    ):
        vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        # todo: (Avihu): Hack! siglip2 naflex now supports pad-free
        # todo: This was done by manually editing the siglip modeling code
        # todo: Until we have a better solution, to run this, consult with me
        # Note! siglip gets a stacked tensor
        image_features = self.vision_tower(pixel_values, spatial_shapes=spatial_shapes,
                                           pixel_attention_mask=pixel_attention_mask, output_hidden_states=True)
        # If we have one vision feature layer, return the corresponding hidden states,
        # otherwise, select the hidden states of each feature layer and concatenate them
        if isinstance(vision_feature_layer, int):
            selected_image_feature = image_features.hidden_states[vision_feature_layer]
        else:
            hs_pool = [image_features.hidden_states[layer_idx] for layer_idx in vision_feature_layer]
            selected_image_feature = torch.cat(hs_pool, dim=-1)

        image_features = self.multi_modal_projector(selected_image_feature)
        # Note (Avihu): downsampling would and newline is more complex at the moment
        assert self.downsampler is None, "downsampler not supported for naflex yet"
        assert self.image_newline is None, "newline not supported for naflex yet"
        return image_features


    
    def get_placeholder_mask(
        self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
        else:
            special_image_mask = input_ids == self.config.image_token_id
            
        n_image_tokens = special_image_mask.sum()
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if inputs_embeds[special_image_mask].numel() != image_features.numel():
            raise ValueError(
                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {image_features.shape[0]}"
            )
        return special_image_mask

    @can_return_tuple
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        image_sizes: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        vision_feature_layer: Optional[Union[int, list[int]]] = None,
        vision_feature_select_strategy: Optional[str] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        spatial_shapes: Optional[torch.LongTensor] = None,
        pixel_attention_mask: Optional[torch.Tensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Union[tuple, Granite4VisionModelOutputWithPast]:
        r"""
        vision_feature_select_strategy (`str`, *optional*, defaults to `"default"`):
            The feature selection strategy used to select the vision feature from the vision backbone.
            Can be one of `"default"` or `"full"`. If `"default"`, the CLS token is removed from the vision features.
            If `"full"`, the full vision features are used.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        vision_feature_layer = (
            vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
        )
        vision_feature_select_strategy = (
            vision_feature_select_strategy
            if vision_feature_select_strategy is not None
            else self.config.vision_feature_select_strategy
        )

        if (input_ids is None) ^ (inputs_embeds is not None):
            print(input_ids, inputs_embeds, position_ids, pixel_values, image_sizes, kwargs, )
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None and pixel_values.size(0) > 0:
            
            if spatial_shapes is not None and pixel_attention_mask is not None:
                # naflex setup
                image_features = self.get_image_features_naflex(
                    pixel_values,
                    spatial_shapes, 
                    pixel_attention_mask,
                    vision_feature_layer=vision_feature_layer
                )
            else:
                image_features = self.get_image_features(
                    pixel_values,
                    image_sizes,
                    vision_feature_layer=vision_feature_layer,
                    vision_feature_select_strategy=vision_feature_select_strategy,
                )
                image_features = torch.cat(image_features, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            special_image_mask = self.get_placeholder_mask(
                input_ids, inputs_embeds=inputs_embeds, image_features=image_features
            )
            inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
        elif torch.is_grad_enabled():
            self.run_dummy_encoder_forward(inputs_embeds, vision_feature_layer, vision_feature_select_strategy)
        try:
            outputs = self.language_model(
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=use_cache,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=True,
                cache_position=cache_position,
                **kwargs,
            )
        except Exception as e:
            print(e)
            print(attention_mask)
            print(position_ids)
            print(inputs_embeds)
            print(input_ids)
            print(kwargs)
            raise e
        return Granite4VisionModelOutputWithPast(
            last_hidden_state=outputs.last_hidden_state,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            image_hidden_states=image_features if pixel_values is not None else None,
        )

    def run_dummy_encoder_forward(self, inputs_embeds, vision_feature_layer, vision_feature_select_strategy):
        if isinstance(self.config.vision_config, Siglip2VisionConfig):
            print("no pixel values, using dummy data to get grads - naflex mode")
            dummy_pixel_values = torch.zeros((1, 256, 768), dtype=inputs_embeds.dtype, device=inputs_embeds.device)
            dummy_spatial_shapes = torch.tensor([[16, 16]], device=inputs_embeds.device)
            dummy_pixel_attention_mask = torch.ones((1,256), device=inputs_embeds.device)
            other_embeds = self.get_image_features_naflex(
                    dummy_pixel_values,
                    dummy_spatial_shapes, 
                    dummy_pixel_attention_mask,
                    vision_feature_layer=vision_feature_layer
                )
            other_embeds = other_embeds[0][:1] * 0 # adding zeros tensor
            inputs_embeds[0, :1] = inputs_embeds[0, :1] + other_embeds

        else:
            print("no pixel values, using dummy data to get grads")
            dummy_data = torch.zeros(
                (3, 3, 384, 384), dtype=inputs_embeds.dtype, device=inputs_embeds.device
            )
            dummy_sizes = torch.tensor([[768, 384]], device=inputs_embeds.device)
            other_embeds = self.get_image_features(dummy_data, dummy_sizes, 
                                                    vision_feature_layer=vision_feature_layer,
                                                    vision_feature_select_strategy=vision_feature_select_strategy)
            other_embeds = other_embeds[0][:1] * 0 # adding zeros tensor
            inputs_embeds[0, :1] = inputs_embeds[0, :1] + other_embeds

class Granite4VisionForConditionalGenerationNaflex(Granite4VisionForConditionalGeneration):
    config_class = Granite4VisionConfigNaflex